#01 · AI agents and MCP · 6.3 minLocal AI Agent in Python: No API Key (Ollama)
Build and run a tool-calling AI agent on your own laptop with Python and Ollama. No API key, no bill.
Python 3.13.2 | ollama==0.6.3191 videos tagged "python tutorial".
#01 · AI agents and MCP · 6.3 minBuild and run a tool-calling AI agent on your own laptop with Python and Ollama. No API key, no bill.
Python 3.13.2 | ollama==0.6.3
#02 · RAG and vector search · 6 minBuild a working RAG system in plain Python with Ollama, and see exactly what it retrieves. No framework, no vector database, no API key.
Python 3.12.15 | ollama==0.6.3
#03 · RAG and vector search · 6.3 minSame notes, same question, three vector databases on your laptop: pgvector, Chroma and Qdrant, with real output and three classic mistakes.
Python 3.12.15 | ollama==0.6.3, psycopg[binary]==3.3.6, pgvector==0.5.1, chromadb==1.5.9, qdrant-client==1.20.0
#05 · LLM engineering and evals · 5.7 minTurn messy emails into structured tickets with a local model and Pydantic, then test the answers. Valid JSON is not the same as correct JSON.
Python 3.12.15 | ollama==0.6.3, pydantic==2.14.1
#06 · LLM engineering and evals · 5.9 minBuild a small LLM evaluation harness with pytest and a local model, then see why six out of six does not mean your prompt is good.
Python 3.12.15 | ollama==0.6.3, pydantic==2.14.1, pytest==9.2.0
#08 · Python for data · 6.1 minRun pandas, Polars and DuckDB on the same ten million rows on one laptop, time them carefully, and check that the answers agree.
Python 3.12.15 | pandas==3.1.0, polars==2.1.0, duckdb==2.0.0, pyarrow==26.0.0
#12 · AI agents and MCP · 6.1 minBuild a pipeline and a router in plain Python with a local model, then test the four lines that guard a supervisor loop, using a scripted stand-in.
Python 3.12.15 | ollama==0.6.3, pydantic==2.14.1
#18 · AI security and governance · 5.6 minI attacked my own language model app and a secret leaked. Then I ran the same attacks against three defences, and in this small test one of them made it worse.
Python 3.12.15 | ollama==0.6.3
#19 · AI security and governance · 6.5 minPatterns found 5 of 8 pieces of personal data, a local model found the rest, and one wrong answer taught me something. A redaction step you can score.
Python 3.12.15 | ollama==0.6.3
#20 · AI coding and productivity · 5.6 minI let an AI fix three bugs: one fix passed its tests, two failed, and every fix changed five lines. The copy, test and diff workflow that caught it.
Python 3.12.15 | ollama==0.6.3, pytest==9.2.0
#22 · AI agents and MCP · 6.7 minOne tool kept failing. One agent loop stopped at 6 calls, one at 10, and one kept going.
Python 3.12.15 | openai-agents==0.24.0, langgraph==1.2.16, langchain-openai==1.7.2, langchain-core==1.7.0, openai==3.36.1
#23 · AI agents and MCP · 6.2 minMy agent forgot my name after four messages. When I let it remember everything, it remembered my card number too.
Python 3.12.15 | ollama==0.6.3
#24 · AI agents and MCP · 4.9 minMy prompt told the agent to ask before deleting. It deleted the file anyway. Here is a gate in code that stopped it.
Python 3.12.15 | ollama==0.6.3
#25 · AI agents and MCP · 5.7 minMy AI tool failed, and the model made up the answer. Here is how to retry and fall back inside the tool.
Python 3.12.15 | ollama==0.6.3
#26 · AI agents and MCP · 5.8 minOne example in the prompt took a small model from 2 of 6 to 6 of 6. Reflection added nothing.
Python 3.12.15 | ollama==0.6.3, pydantic==2.14.1
#27 · AI agents and MCP · 5.7 minMy agent loop could run forever. Here are the three limits that stop it, and one thing I did not expect from a real model.
Python 3.12.15 | ollama==0.6.3
#28 · AI agents and MCP · 5.3 minI gave two MCP servers a tool called search each. Merge their tools by name, and one of them vanishes.
Python 3.12.15 | mcp==2.3.0, ollama==0.6.3
#29 · AI agents and MCP · 5.3 minTools are only one of three things an MCP server can offer. Here are the other two, and the one default that stopped a file leak.
Python 3.12.15 | mcp==2.3.0
#30 · AI agents and MCP · 5.2 minAn MCP server on a network with no check answers anyone who can reach it, so I added a token check.
Python 3.12.15 | mcp==2.3.0
#31 · AI agents and MCP · 6.1 minA tool on my MCP server failed with an error a model could do nothing with. The Inspector's command line mode showed me that error exactly as a model would get it.
Python 3.12.15 | mcp==2.3.0
#32 · AI agents and MCP · 5.5 minTwo agents that share no code can still work together, if they share a protocol.
Python 3.12.15 | a2a-sdk[http-server]==1.2.3, uvicorn==0.54.0
#33 · AI agents and MCP · 6.4 minA browser script tied to one page's ids can break the day the page changes. On my two test pages, a loop that looks first did not.
Python 3.12.15 | playwright==1.64.1
#34 · AI agents and MCP · 5.5 minNo attack I tried reached another customer's order. I still moved the customer check into the code.
Python 3.12.15 | ollama==0.6.3
#35 · AI agents and MCP · 5.1 minWhich step of my agent failed, and which was slow? One trace showed both, with the time each step took.
Python 3.12.15 | opentelemetry-sdk==1.45.1
#36 · AI agents and MCP · 5.3 minTwo agents argued, and did no better than a plain vote, for more calls.
Python 3.12.15 | ollama==0.6.3
#37 · AI agents and MCP · 5.7 minA read only database stopped three of eight attacks on my SQL agent. The other five needed other layers.
Python 3.12.15 | duckdb==2.0.0
#38 · AI agents and MCP · 5.2 minMy job crashed, I restarted it, and it sent the same email twice.
Python 3.12.15
#39 · AI agents and MCP · 5.2 minI cut my chat to the last three turns and the model forgot all five facts.
Python 3.12.15 | ollama==0.6.3
#40 · RAG and vector search · 5.5 minChunking decides what your RAG app can find. I cut one document five ways and scored each against ten questions.
Python 3.12.15 | ollama==0.6.3
#41 · RAG and vector search · 5.7 minKeyword search and vector search fail in opposite ways. I merged them and scored all three on twelve questions.
Python 3.12.15 | ollama==0.6.3, rank_bm25==0.2.2, numpy==2.5.3
#42 · RAG and vector search · 4.9 minMy search found the right page but did not put it first. Can a reranker fix that?
Python 3.12.15 | fastembed==0.9.1
#43 · RAG and vector search · 5.7 minAn early version of my RAG check passed the answer 'I don't know'. The fixed version still gets things wrong.
Python 3.12.15 | ollama==0.6.3
#44 · RAG and vector search · 5.4 minOne page of my PDF came back with no text of its own, and nothing warned me.
Python 3.12.15 | pypdf==6.21.4, fpdf2==2.8.9, pdfplumber==0.11.10, pillow==13.0.0
#45 · RAG and vector search · 4.8 minMy search put the right page first 7 times out of 12. A pretend answer did it 10 times.
Python 3.12.15 | ollama==0.6.3
#46 · RAG and vector search · 5.3 minWith two sentences, text search had the facts for 1 of 6 questions. A graph written by hand had them for all 6.
Python 3.12.15 | ollama==0.6.3
#47 · RAG and vector search · 5.9 minI tried to break a RAG app six ways. Four broke, one barely, and one did not break in five tries.
Python 3.12.15 | ollama==0.6.3
#48 · RAG and vector search · 4.9 minThe smallest embedding model got 16 of 30. Three others got all 30, and one is a third the size.
Python 3.12.15 | fastembed==0.9.1
#49 · RAG and vector search · 5.6 minI searched 100,000 vectors by looking at about 500, and still found nearly every right answer.
Python 3.12.15 | faiss-cpu==1.15.0
#50 · RAG and vector search · 5.8 minOn a table built as a hard case, a filtered vector search returned no rows, no error, and the plan said the index was used.
Python 3.12.15 | psycopg[binary]==3.3.6, pgvector==0.5.1, numpy==2.5.4
#51 · RAG and vector search · 6.2 minSame filter, same users: applied after the search it found the user's own note 18 times out of 36; applied before the search, all 36.
Python 3.12.15 | ollama==0.6.3
#52 · RAG and vector search · 5.7 minEvery citation was real and every quote exact. With the right source removed, two still pointed at the wrong source.
Python 3.12.15 | ollama==0.6.3
#53 · RAG and vector search · 5.4 minWith a loose threshold, a semantic cache answered 'Do you ship to Norway?' with the answer for Ireland.
Python 3.12.15 | ollama==0.6.3
#54 · RAG and vector search · 5.4 minPasting everything into a 4,096 token prompt got four of four right, until the catalogue doubled.
Python 3.12.15 | ollama==0.6.3
#55 · RAG and vector search · 5 minI searched twelve drawn pictures with sentences. It found the blue square. Left of a square, it got wrong.
Python 3.12.15 | fastembed==0.9.1
#56 · LLM engineering and evals · 5.3 minI changed a prompt to make it better and the score went down. A score alone could not tell me which cases broke.
Python 3.12.15 | ollama==0.6.3
#57 · LLM engineering and evals · 5 minI swapped the order of two answers and my judge changed its mind on half the questions in one test.
Python 3.12.15 | ollama==0.6.3
#58 · LLM engineering and evals · 5.2 minJSON mode with no field names in the prompt fit my fields 0 times out of 12. With the keys named: 12 of 12.
Python 3.12.15 | ollama==0.6.3, pydantic==2.14.1
#59 · LLM engineering and evals · 4.6 minThe same prompt showed its first word after 0.2 seconds instead of eight.
Python 3.12.15 | ollama==0.6.3
#60 · LLM engineering and evals · 5 minIn a test with injected failures, retries fixed a simulated busy server, then made twelve attempts and fixed nothing.
Python 3.12.15 | ollama==0.6.3
#61 · LLM engineering and evals · 5.3 minI sent 1,086 tokens into a 1,024-token context on purpose. No error, and the first rule was not followed.
Python 3.12.15 | ollama==0.6.3
#62 · LLM engineering and evals · 5.4 minI ran the same question 20 times under six settings. Two gave 20 different answers. Four gave exactly one.
Python 3.12.15 | ollama==0.6.3
#63 · LLM engineering and evals · 5.6 minZero examples scored 12 of 20. Eight scored 18. Flipped labels dropped it to 13.
Python 3.12.15 | ollama==0.6.3
#64 · LLM engineering and evals · 5.1 minOn five multi-step problems my model got 0 right answering directly and 4 right thinking step by step.
Python 3.12.15 | ollama==0.6.3A made-up change dropped the pass rate from 90 to 60 percent, and the pytest gate exited 1.
Python 3.12.15 | pytest==9.1.1A small model trained on 12 hand labels got 149 of 200 right; trained on a teacher's labels, 188.
Python 3.12.15
#68 · LLM engineering and evals · 5 minI asked a small model for 27 test questions and only 17 passed three cheap checks.
Python 3.12.15 | ollama==0.6.3
#69 · LLM engineering and evals · 6.1 minOne check caught 1 of 8 made up answers. A strict model judge flagged 6 of 8 good ones.
Python 3.12.15 | ollama==0.6.3
#70 · LLM engineering and evals · 5.4 minA one-letter typo in a prompt template renders as nothing, and nothing warns you.
Python 3.12.15 | jinja2==3.1.6Routing cost 19.1 made-up cents instead of 40, and got one more of 20 questions wrong.
Python 3.12.15Four bits in blocks of 32 lost about 11 percent; four bits with one scale lost almost 98.
Python 3.12.15
#83 · Local and open models · 5.7 minReading a 2,542 token prompt took my laptop 155 seconds.
Python 3.12.15 | ollama==0.6.3
#87 · Local and open models · 5.5 minOne model, one laptop: 1.6, 6.6 and 9.2 tokens per second. All three were real.
Python 3.12.15 | ollama==0.6.3
#88 · AI security and governance · 5.3 minA sentence I planted in a made-up note told my assistant what to say, and it said it.
Python 3.12.15 | ollama==0.6.3
#89 · AI security and governance · 5.4 minMy bot refused a plain question, then answered four of eight wrapped versions of it.
Python 3.12.15 | ollama==0.6.3
#90 · AI security and governance · 5 minI hid six fake keys in a project. My first scanner found three.
Python 3.12.15
#91 · AI security and governance · 5.3 minIn my test, one simulated user sent 400 big requests in 20 seconds, and my API answered all of them.
Python 3.12.15 | fastapi==0.161.1, starlette==1.9.0, httpx==0.28.1
#92 · AI security and governance · 5.2 minMy demo notes agent could delete every note it held, and nothing in the code would stop it.
Python 3.12.15 | ollama==0.6.3
#93 · AI security and governance · 5.5 minA planted page reached the top three results for four of five questions, and a filter missed two leaks.
Python 3.12.15 | ollama==0.6.3
#94 · AI security and governance · 4.9 minI told my bot never to reveal its instructions, put a code in them, asked eight ways, and five worked.
Python 3.12.15 | ollama==0.6.3
#95 · AI security and governance · 5 minAll ten of my model replies passed every check, and three were still wrong.
Python 3.12.15 | ollama==0.6.3
#96 · AI security and governance · 5 minMy banned-word list blocked a made-up test comment about killing a Python process.
Python 3.12.15 | ollama==0.6.3
#97 · AI security and governance · 5.4 minBefore you read the EU AI Act, one question decides most of it: which tier is your app in?
Python 3.12.15
#98 · AI security and governance · 5 minI edited one line of my own audit log. It noticed. Then I edited the last line.
Python 3.12.15
#99 · AI security and governance · 5.3 minLoading a pickle file, a format model files can use, ran code I had hidden inside it, four files out of four.
Python 3.12.15 | safetensors==0.8.0, numpy==2.5.4
#100 · AI security and governance · 5.2 minA quick comma-split loader failed on all 200 rows of realistic fake customers.
Python 3.12.15 | Faker==40.41.0
#101 · AI security and governance · 5.4 minA made-up count of 412 told a simulated attacker whether one person was in the data.
Python 3.12.15
#102 · AI security and governance · 5 minFour of six test attacks worked on my demo agent; two were plain delete and send requests.
Python 3.12.15 | ollama==0.6.3On 50,000 made-up orders, a Parquet file was 52% of the CSV, and one query needed the amount column from 1 row group of 5.
Python 3.12.15 | pyarrow==26.0.0On 60,000 made-up orders, a one-month query opened 1 of 12 files by month, but 300 of 300 by customer.
Python 3.12.15 | pyarrow==26.0.0With the same simulated crash at offset 2, commit-after handled order 3 twice and commit-before lost it.
Python 3.12.15B was ahead with p 0.015; stopping at the first of ten looks raised false alarms from 4.9 to 17.2 percent.
Python 3.12.15
#137 · Python for data · 6.3 minOne made-up report in pandas and Polars: same 4 rows, South at 6,814.5 in both.
Python 3.12.15 | pandas==3.1.0, polars==2.1.0
#138 · Python for data · 6.5 minSeven pandas changes on 1,000,000 made-up rows: four cleared a 3x speed bar, one did not, two cut memory.
Python 3.12.15 | numpy==2.5.4, pandas==3.1.0
#139 · Python for data · 6.8 minSix made-up CSV rows through Pydantic: 3 valid, 3 rejected, with 6 named errors in one pass.
Python 3.12.15 | annotated-types==0.8.0, pydantic==2.14.1, pydantic-core==2.50.1, typing-extensions==4.16.0, typing-inspection==0.4.4
#140 · Python for data · 6.7 minAdd one day to a London time across the clock change, and only 23 real hours pass.
Python 3.12.15 | tzdata==2026.5On Python 3.12, the built-in robots.txt parser and the RFC 9309 longest-match rule differed on 2 of 8 rows.
Python 3.12.15
#142 · Python for data · 6.7 minA local API fails 3 times on purpose, and the reliable client still gets all 23 items in 8 requests.
Python 3.12.15Mypy reported 4 errors from 3 bugs without running the code, and missed a wrong average.
Python 3.12.15 | mypy==2.4.0
#146 · Python for data · 6 minSix pytest tests pass, and one catches a column that came back float64 instead of int64.
Python 3.12.15 | numpy==2.5.4, pandas==3.1.0, pytest==9.1.1
#149 · Python for data · 6.4 minOn made-up data, dropping rows moved a mean from 40,057 to 37,458; a model on age got back to 39,972.
Python 3.12.15 | numpy==2.5.4, pandas==3.1.0, python-dateutil==3.0.0, six==1.17.0, tzdata==2026.5
#150 · Python for data · 6 minRead all at once, a 4 MB made-up CSV peaked at over 10x its size; streamed, under 1 MB.
Python 3.12.15
#151 · Python for data · 6.6 minSeven made-up messy notes: two regex patterns pull out 5 order ids and 5 phone numbers, and miss 1.
Python 3.12.15Version 3 scored 0.867 against 0.897 in production: the registry refused it, a forced push went through, and one call rolled it back.
Python 3.12.15Moving the mean up by 4 gave a PSI of 0.153; the K-S test flagged a shift of 1 that PSI scored as small.
Python 3.12.15On made-up data, a naive join scored a churn rule at 0.96; point in time, the same rule scored 0.69.
Python 3.12.15In a made-up day, a 40-cent user cap and a 60-cent daily cap each refused a call before it ran.
Python 3.12.15I planted 5 made-up secrets in a made-up ML app; a two-rule scanner found 4, missed 1 and raised 1 false alarm.
Python 3.12.15In a made-up simulation, batches of 256 processed 481 items a second against 45 one at a time.
Python 3.12.15
#169 · AI coding and productivity · 6.4 minTwo of my seven rules could not be checked; a script checks the other five.
Python 3.12.15
#171 · AI coding and productivity · 6.4 minA first draft passed all five tests and still charged minus two pounds.
Python 3.12.15
#172 · AI coding and productivity · 5.2 minA tidy refactor changed the price of every heavy express parcel. Sixty four recorded cases caught it.
Python 3.12.15 | pytest==9.1.1
#173 · AI coding and productivity · 6.6 minEight rows crashed a report; a small shrink found the one row and a 5-line repro.
Python 3.12.15
#174 · AI coding and productivity · 5.4 minA seven point checklist found seven of eight planted problems in AI style code. A security linter found three.
Python 3.12.15 | pytest==9.1.1, bandit==1.9.4
#177 · AI coding and productivity · 6.4 minAn audit that runs a quick sign-up app finds 4 problems; the fixed version passes all 4 checks.
Python 3.12.15
#179 · AI coding and productivity · 6.6 minThree one-click resolutions left no conflict markers, and each one still failed a test.
Python 3.12.15
#180 · AI coding and productivity · 6.3 minA quick first try passed 1 of 7 spec rules; code written to the spec passed 7 of 7.
Python 3.12.15
#182 · Machine learning and data science · 6.2 minBy hand and with scikit-learn, the slope came out at 5.77 both times. The true slope was 5.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#183 · Machine learning and data science · 6.5 minA model that never flags fraud scored 98% accuracy and caught 0 of 20 frauds.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#184 · Machine learning and data science · 6 minDegree 14 had a training error of 0.00 and a validation error of 3.27.
Python 3.12.15 | numpy==2.5.4
#185 · Machine learning and data science · 6.3 minOn pure noise with random labels, cross-validation reported 0.89 accuracy, until one step moved.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#186 · Machine learning and data science · 6.2 minOne column filled in after delivery lifted test accuracy from 0.69 to 0.95.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#187 · Machine learning and data science · 6.2 minOn made-up rents, a straight line with one extra column missed by 31.2 a month; a 500-tree forest on the raw columns missed by 39.4.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#189 · Machine learning and data science · 6.3 minOne deep tree scored 0.728 on test; a hundred random trees together scored 0.842.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#192 · Machine learning and data science · 6.2 minYesterday as the forecast missed by 12.3 a day; the same weekday last week missed by 6.8, and a model only got to 5.4.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#193 · Machine learning and data science · 6.2 mink-means matched 299 of 300 points on round clouds, but turned a small cloud of 40 into a cluster of 102.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#194 · Machine learning and data science · 6.4 minA z-score found 1 of 3 planted bad readings; a median-based check found all 3.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#195 · Machine learning and data science · 6.3 minA program learned three numbers for each of 24 made-up films, and films of one genre ended up together in 3D.
Python 3.12.15On 200 made-up points, a hidden layer took training accuracy from 0.59 to 0.99, and a gradient check caught a bug.
Python 3.12.15 | numpy==2.5.4
#199 · Machine learning and data science · 6.3 minA 95% bootstrap interval from 10 made-up values caught the true mean 863 times in 1000.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#200 · Machine learning and data science · 6.6 minA plain model caught 8 of 39 made-up frauds; class weights, a tuned threshold and resampling each caught 30 to 34.
Python 3.12.15 | numpy==2.5.4, scikit-learn==1.9.1, scipy==1.18.1, joblib==1.6.0, narwhals==2.27.1, threadpoolctl==3.7.0, cloudpickle==3.1.2
#203 · Data governance · 5.6 minA glossary in a file, and four checks that find its four faults in a second.
Python 3.12.15 | PyYAML==6.0.3
#204 · Data governance · 6.1 minA 25-line PII scanner tagged email, phone and postcode right, and order references wrong.
Python 3.12.15
#205 · Data governance · 5.7 minWhich columns feed this number? SQLGlot can answer in 15 lines, and I will show you where it misleads.
Python 3.12.15 | sqlglot==30.23.0
#208 · Data governance · 5.8 minI erased a person, a scan found her still in an export, and a restore brought her back.
Python 3.12.15
#209 · Data governance · 5.2 minA UK GDPR data map in a file, and a 15-line check that found 3 of 5 activities to fix.
Python 3.12.15
#210 · Data governance · 5.3 minOrders averaged 97.7 percent, amber. One critical check made it red.
Python 3.12.15
#211 · Data governance · 5.9 minVersion 1.3.0 removed a column and changed a type. The check said it needs 2.0.0.
Python 3.12.15 | PyYAML==6.0.3
#212 · Data governance · 5.3 minA five-line file in git says who owns each dataset. A script found a gap and a ghost team.
Python 3.12.15
#213 · Data governance · 5.5 minTen records, five used. A dataset card says why the other five were left out.
Python 3.12.15
#214 · Data governance · 5.8 minWe listed seven AI systems. A short script found two more nobody had listed. Here is the register.
Python 3.12.15 | PyYAML==6.0.3
#216 · Data governance · 6.5 minSix made-up pipeline runs, three gaps: an evidence report built straight from the logs.
Python 3.12.15
#219 · Data strategy · 5.4 minWhich AI project first? A scoring model picks a winner. I tested whether the winner survives a different opinion.
Python 3.12.15
#227 · Data strategy · 5.1 minIn a made-up traffic mix, one kind of query was 2 percent of the traffic and 60 percent of the compute cost. Here is how to find it.
Python 3.12.15 | duckdb==2.0.0
#229 · Enterprise architecture and AI agents · 5.5 minI wrote a made-up agent platform down as data, planted three bad arrows among sixteen, and a check found all three.
Python 3.12.15
#230 · Enterprise architecture and AI agents · 5.6 minI sent eight requests through one gateway, and five were stopped at the door.
Python 3.12.15 | fastapi==0.161.1, starlette==1.9.0, httpx==0.28.1
#231 · Enterprise architecture and AI agents · 5.5 minSeven tool calls worked directly. Through a small gateway, four were blocked.
Python 3.12.15 | mcp==2.3.0
#232 · Enterprise architecture and AI agents · 5.4 minYour agents all share one API key. When one overwrites a file, the log cannot say which. Here is the fix.
Python 3.12.15 | PyJWT==2.15.1
#233 · Enterprise architecture and AI agents · 5.9 minI tried eight ways to use an agent's token, and only the allowed one worked.
Python 3.12.15
#234 · Enterprise architecture and AI agents · 5.6 minI changed one line of policy text, and a fifty pound refund was denied, with the checking code unchanged.
Python 3.12.15 | cedarpy==4.12.1
#235 · Enterprise architecture and AI agents · 5.6 minOne line in my log had no request id, and it was the audit record of a refund.
Python 3.12.15
#236 · Enterprise architecture and AI agents · 5.3 minThe same three agents solved four tickets, or six, depending on how I organised them.
Python 3.12.15
#237 · Enterprise architecture and AI agents · 5.8 minI killed my agent in the middle of a write, then reloaded its memory. Three kinds of memory behaved in three different ways.
Python 3.12.15
#238 · Enterprise architecture and AI agents · 5.6 minAfter a crash, my toy agent wrote two replies to the same ticket. The agent was fine.
Python 3.12.15
#239 · Enterprise architecture and AI agents · 5.4 minIn a simulation, a client that retried on a timeout nearly doubled the agent's work, and got half the answers.
Python 3.12.15
#240 · Enterprise architecture and AI agents · 5.8 minSeven operations in the old API, four tools for the agent, and only two of six test calls got through.
Python 3.12.15
#241 · Enterprise architecture and AI agents · 5.8 minA diagram can hide a shortcut your own rules do not allow. Check the model before you draw it.
Python 3.12.15
#242 · Enterprise architecture and AI agents · 6 minI planted seven problems in seven decision records for an AI assistant, and a small lint found all seven.
Python 3.12.15
#243 · Enterprise architecture and AI agents · 5.6 minIn a simulation, one busy agent left the support agent with one call served out of twenty.
Python 3.12.15
#244 · Enterprise architecture and AI agents · 5.5 minOne agent in a loop used 88 percent of a made-up day's spend, and nothing stopped it.
Python 3.12.15
#245 · Enterprise architecture and AI agents · 5.8 minWhen an agent gives a bad answer, the hard part is getting that exact run back.
Python 3.12.15 | opentelemetry-api==1.45.1, opentelemetry-sdk==1.45.1
#246 · Enterprise architecture and AI agents · 5.6 minI asked eight questions to three stores, and each got only its own kind right.
Python 3.12.15
#247 · Enterprise architecture and AI agents · 5.6 minI gave every tenant its own database, and one tenant's data still leaked.
Python 3.12.15
#248 · Enterprise architecture and AI agents · 6.1 minWhere does an AI agent go on an enterprise architecture diagram?
Python 3.12.15
#249 · Enterprise architecture and AI agents · 6.1 minI built an agent platform in 100 lines, and switched an agent off by changing one flag in the registry.
Python 3.12.15Asked one question, a one-table design I built without tenant, version or access columns returned 3 chunks in its top 4 that the user should not get: an old version, another tenant's text, and a note she may not read.
Python 3.12.15On one made-up day, a wide table I wrote to fail could not find the busiest hour, overstated one driver's earnings by 54.18, and gave a surge of 1.0 where the schema said 1.8.
Python 3.12.15
#274 · How it works, by example · 6.4 minHuffman coding took a 102 character text from 816 bits to 349, before its code table, in about 40 lines of Python.
Python 3.12.15
#275 · How it works, by example · 6.5 minA 10,000 bit Bloom filter gave 0.9% false positives, then 41.6% once I overfilled it.
Python 3.12.15
#278 · How it works, by example · 6.4 minA cache of 100 keys out of about 1,000 answered 58% of made-up requests.
Python 3.12.15
#281 · How it works, by example · 6.4 minTwo users, one password, the same plain hash. A salt and a slow hash fix it.
Python 3.12.15
#282 · How it works, by example · 6.4 minBrute force: 6 of 6 guesses through, then 2 of 6 once the lockout was on.
Python 3.12.15
#292 · AI agents and MCP · 5.7 minPeople talk about MCP like magic, so I spoke it by hand, in plain JSON lines.
Python 3.12.15 | mcp==2.3.0
#293 · RAG and vector search · 5.2 minI asked a model five questions about a cafe, and it got none right, until I gave it the facts.
Python 3.12.15 | ollama==0.6.3
#296 · LLM engineering and evals · 6.4 minWith hand-set weights, the word tired gave 0.79 of its attention to cat, in about 40 lines of NumPy.
Python 3.12.15 | numpy==2.5.4
#297 · SQL and analytics · 6.8 minTen SQL queries on one small shop: a left join finds the 1 customer who never ordered.
Python 3.12.15
#301 · RAG and vector search · 7 minSix kinds of RAG, one small test: the plainest one scored 10 of 12, and the graph one answered a question none of the others could.
Python 3.12.15 | fastembed==0.9.1, numpy==2.5.4
#302 · Forward deployed engineer interviews · 6 minA list of keywords scored as high as a real interview answer. Here is what told them apart.
Python 3.12.15
#303 · Forward deployed engineer interviews · 6.4 minThis plan sounded fine. A 53-line checker found eight holes in it.
Python 3.12.15
#304 · Forward deployed engineer interviews · 5.9 minBoth pilots were faster. One still failed a guardrail that was written before the results.
Python 3.12.15
#305 · Forward deployed engineer interviews · 5.7 minA plain retry loop put five duplicate orders into a stand-in system. One key fixed it.
Python 3.12.15
#306 · Forward deployed engineer interviews · 6.1 minAnyone can POST to your webhook URL. Four questions to ask of every message: forged, changed, old or repeated?
Python 3.12.15
#307 · Forward deployed engineer interviews · 6.2 minA guessed fix for a changed export made two totals 100 times too big, with no error.
Python 3.12.15
#308 · Forward deployed engineer interviews · 5.8 minIn a simulation, last write wins silently lost 340 of 1,000 edits. Two other ways lost none.
Python 3.12.15
#309 · Forward deployed engineer interviews · 6.3 minMy data profile said 275 blank amounts. The file had 363. Here is why.
Python 3.12.15 | duckdb==2.0.0
#310 · Forward deployed engineer interviews · 5.9 minThree of seven work orders were quietly wrong. A 60-line model found them and refused a bad change.
Python 3.12.15
#312 · Forward deployed engineer interviews · 5.7 minA swapped package installed with no error. With hash checking on, it was refused.
Python 3.12.15 | pip==26.2.2
#314 · Forward deployed engineer interviews · 5.9 minCustomer B got customer A's webhook address and secret reference. The bug was one line.
Python 3.12.15
#315 · Forward deployed engineer interviews · 5.3 minThe same bug reached 100, 30 or 830 requests. The difference was a canary that someone watched.
Python 3.12.15
#316 · Forward deployed engineer interviews · 6 minA model that only remembers scored 100% on my test set. Split another way, 40%.
Python 3.12.15
#317 · Forward deployed engineer interviews · 6 minVersion 2 scored 80% against 70%. On 40 cases, I still could not tell them apart.
Python 3.12.15
#318 · Forward deployed engineer interviews · 5.9 minCutting the prompt barely helped. A small model with shorter answers made it fit.
Python 3.12.15
#319 · Forward deployed engineer interviews · 6.2 minThe biggest cause of failures was hiding in a bucket called 'other'.
Python 3.12.15
#320 · Forward deployed engineer interviews · 6 minThe alert fired at 10:09. The cause was at 10:02, in a different system.
Python 3.12.15
#321 · Forward deployed engineer interviews · 5.9 minThe obvious redaction rule found nothing, and five secrets were still in the support bundle.
Python 3.12.15
#322 · Forward deployed engineer interviews · 5.5 minBoth promises were kept. One afternoon burned the budget at 57 times the allowed rate.
Python 3.12.15
#323 · Forward deployed engineer interviews · 5.9 minThe plan said 34 days. One simulated project in a hundred finished that fast.
Python 3.12.15
#324 · Forward deployed engineer interviews · 6 minThree copied scripts, 33 repeated lines. One 16-line connector gave the same orders.
Python 3.12.15
#325 · Forward deployed engineer interviews · 6.5 minA messy file, five checks, five attempts. The first passes 1 of 5. The last passes all 5.
Python 3.12.15
#329 · Forward deployed engineer interviews · 5.3 minConfigure won 40 to 34. Moving one weight at a time shows how sure that is.
Python 3.12.15
#330 · AI agents and MCP · 6.8 minA plain keyword router put the right tool in a list of 5 for 28 of 30 made-up requests.
Python 3.12.15One toy, two dates: the same MCP tool said 3.16, then 3.17, because the date picked the schedule.
Python 3.12.15 | mcp==2.3.0