AI

Embeddings

I built AI-powered product search systems using embeddings.

See its threads in the world

Used on

Stories where it mattered

2021 · Hybrid.Chat conversational platform

High availability for a bot runtime that insisted on being in memory

Situation
The chatbot platform's runtime holds conversation state in memory, so under high availability any pod restart dropped live conversations.
Task
Survive pod churn without rewriting the conversational runtime the product was built on.
Action
I checked alternative frameworks, found the same limit, and kept the runtime: sticky sessions, state persisted to Mongo, and Redis pub/sub to close duplicates.
Result
The platform ran multi-container with state surviving restarts, and the design documented its failure modes, including those the fix introduced.
View in the world
2023 · Hybrid.Chat conversational platform

Wanting to build it in house and buying instead

Situation
The bot platform needed web push notifications, and my instinct was to build it in house for future flexibility.
Task
Decide between building and buying, separating engineering reasons from personal preference.
Action
I wrote down what I wanted from in-house, weighed it against what the release needed, and recorded the preference and the decision against it.
Result
We shipped on the vendor, and the recorded trade-off made later customisation questions easy to answer.
View in the world
2023 · Omnichannel agent inbox

Taking a component out of the stack instead of adding one

Situation
A messaging product ran Keycloak for identity, yet users, roles and organisations already lived in Dgraph, so every write landed twice.
Task
Remove the dual write and its compensating code, and get to one store of record.
Action
I checked which Keycloak features were used, found none beyond authentication, which NestJS covers, and wrote a decision record proposing its removal.
Result
The dependency and its rollback logic came out, leaving the backend with a single store of record.
View in the world
2024 · Kray.AI search and document assistant

One model for everyone, or one per customer

Situation
A document assistant for a fleet operator needed an architecture, and one model trained on every vessel's documents was the assumed default.
Task
Lay out the real trade-offs of cost, latency and access control instead of just picking one.
Action
I wrote both options out for client and team: one shared model versus one model per vessel, with pros, cons and the replica arithmetic.
Result
The partitioned design went forward, with the loss of cross-fleet questions accepted openly.
View in the world
2026 · Regulatory medical-writing platform

A cache that remembered the failures

Situation
On a regulatory document platform, generated documents came out wrong in ways that did not reproduce, and retries repeated the bad output.
Task
Find the true cause in a pipeline whose output reaches a client in a regulated industry.
Action
I traced it to a global result cache that also stored partial failed runs, then re-architected generation around per-step isolation with failure-aware cleanup.
Result
Failed runs stopped affecting later ones, a retry became a real retry, and the written root cause let the team review it.
View in the world
2026 · Regulatory medical-writing platform

Tables that arrived in the client's document as raw markup

Situation
Medical writing deliverables showed misplaced tables, detached titles, and sometimes raw markup visible as text in the client's document.
Task
Produce one root cause analysis for three separately reported bugs before anyone wrote a fix.
Action
I found three causes: raw XML mutation, incomplete model-emitted HTML, and context-blind placeholders. I moved tables to a structured contract with neighbour context.
Result
The whole class of defect went away, not just the three reported instances.
View in the world
2026 · TalkingDB document question-answering

Credentials in git history, and rotating keys was not enough

Situation
Reviewing repositories before a host move, I found OAuth client secrets in git history, which ships to anyone who clones.
Task
Make the secrets unreachable, not merely invalid, before the repositories were pushed to their new home.
Action
I cloned bare mirrors, mapped each leaked value to a placeholder, rewrote history with filter-repo, pushed to the new remotes, and documented the exact procedure.
Result
The repositories moved with secrets removed from every commit, and another file with hardcoded credentials was flagged for cleanup.
View in the world
2026 · TalkingDB document question-answering

Holding back an accuracy number that would have looked great

Situation
A large reported jump in search accuracy was circulating and was about to become a go-to-market claim.
Task
Check whether the gain was real, since repeated passing results seemed to be counted again across runs.
Action
I raised it plainly, then set a general standard: proposed improvements need a frequency table of failing cases and expected fix share.
Result
The inflated figure never reached a customer-facing claim, and benchmark results are now reviewed against a sheet before approval.
View in the world
2026 · TalkingDB document question-answering

Being asked to pull a date forward, and naming the real constraint

Situation
Leadership wanted a cross-product integration delivered sooner because it strengthened the story with channel partners, which I agreed with.
Task
Show that the date was set by available engineering time, not technical complexity.
Action
I agreed with the goal, offered to make time visible through tracked tasks, and named three ways to buy time: delegation, clear ownership, hiring.
Result
The conversation moved from when it can be done to which trade-offs we accept, with the task breakdown as shared evidence.
View in the world
2026 · TalkingDB document question-answering

A port that turned out not to be a port

Situation
Search training was to be ported from an older e-commerce AI codebase, and the files looked like a simple lift and shift.
Task
Scope the port properly before committing a sequence of stories others would plan around.
Action
I read the code itself: the crawler was tied to one storefront, so I called for a fresh generic crawl-and-chunk pipeline and resequenced.
Result
The order of work changed without changing the plan, and client-specific code stayed out of the generic package.
View in the world
2024 · Regulatory medical-writing platform

Asking what onboarding was supposed to produce

Situation
Two engineers were joining a document platform module covering audit logs, automatic mapping and server management.
Task
Give onboarding an intended outcome rather than just an activity.
Action
I asked the lead for the target outcome, how we would judge pace, and when production value was expected, and offered a reference checklist.
Result
Onboarding ran against a stated target and a first-value date, and the checklist was reused.
View in the world

Goes with

Vector SearchLLMsRAGNLP / NERPythonDocument Intelligence

Throw me a paper plane

Write a line; I’ll find it in my inbox.

✉ Email▤ Résumé☎ Call✆ WhatsAppLinkedIn · GitHub · QR in the world