Notes

What we measured, what failed, and what we chose — written so you can repeat it. Every number comes with the setup it was measured on.

FOLLOW NEW NOTES

RSS feed Ask for them by email

Use the feed in any RSS reader, or write to contact@bodhamlabs.com and we will email you new notes.

LATEST · BENCHMARK · 6 MINVector search on an Android phone: 365,000 chunks in 7 msWe estimated a plain Kotlin loop would need 200–375 ms. On a mid-range phone it took 20 ms on one thread and 7 ms on four. The test, the numbers, and what they do not prove.Read the note

ALL NOTES · 8 PUBLISHED

BENCHMARKVector search on an Android phone: 365,000 chunks in 7 msOur estimate was 10–18× too pessimistic. The test and the caveats.6 min FIELD NOTEWhy our agent memory stopped waiting for a language modelWe took answer-writing out of the query path. Then the memory got used: 111 queries, March to May 2026.5 min FIELD NOTEOur autonomous agent went dark for 45 hours. These are the rules we wrote after.One billing error, a five-link chain, and six guardrails for anyone deploying agents.5 min FIELD NOTEOur Mamba model's hidden state reached 10¹⁵. Normalise your inputs.Unscaled inputs, three dead features, and a GRU whose state never moved.5 min FIELD NOTE2-second stalls from one synchronous disk writeFrom 2,075 ms spikes to a 111 ms maximum, and a WSL networking fix.5 min FIELD NOTEThe pipeline ran 24 hours without a crash. The model still failed.100 Hz, zero crashes, and a model trained on 1.4 seconds of data.5 min FIELD NOTEFrom web developer to an applied-AI studio: how Bodham Labs startedWebsites, payments, a classroom in Delhi, and a year of building AI in the open.4 min GUIDETranslating into Hinglish with an LLM: the rules that stopped the driftVersion 1 started explaining the text instead of translating it. The rules that fixed it.4 min

COMING NEXT · 8

Planned notes, not yet published. Each one becomes a link here when it goes live.

  • BENCHMARKRouting Hinglish speech: comparing language detectorsWhich language detector sent the fewest clips to the wrong model.PLANNED
  • GUIDEFee reminders on WhatsApp for a coaching institute, without getting bannedSending hours, pacing, templates and the official API.PLANNED
  • BENCHMARKSpeech-to-text for Hinglish on Android: comparing models on our own recordingsError rates, memory and speed on a mid-range phone.PLANNED
  • GUIDEGiving AI coding agents long-term memory with MCPHow to build a local memory server, step by step.PLANNED
  • FIELD NOTESelf-healing for a WhatsApp agent on one office PCWhat broke in the first weeks, and what fixed it.PLANNED
  • BENCHMARKSQLite full-text search on a phone: why a postings cap mattersWhy the worst queries were slow, and how a cap fixed them.PLANNED
  • TEACHINGTeaching AI to BCA students: what works in the first weeksNotes from our own classroom.PLANNED
  • GUIDEAI automation for a small business in India: where to startPick one job, measure it, then decide.PLANNED

Want us to test something? Send us the question — if it is useful to others, we will write it up.