The pod is the new firm. Building Manthan — the aOS for it. Ex-Citi/MS IB London. Exited fintech founder. Vertical AI since 2017.
London
Joined August 2011
- Tweets4.3K
- Following1.6K
- Followers796
- Likes5K
Pinned Tweet
1/ The 1000x Investor: what happens when you build an investment fund that never stops thinking?
We built Manthan Intelligence — the AI-native engine for the upcoming Tavaga VC Fund. It analyses 25 companies/day, scores its own predictions, and rewires its own lenses.
1/ We almost deleted 907 real company websites yesterday. The "obvious" fix for a duplicate-records cleanup turned out to be based on a false assumption. Here's what caught it before it shipped.
5/ What we did instead: field-union merge, not delete. Keep the richer, more recent value on every field, every pair. Result: 3 independent precision checks (n=2,485 / 89 / 30) each scored 1.0000. Read-back against the pre-merge snapshot: 0 data loss. 15,500 to 14,442 records, verified.
1/ A third-grader just scored 5 on AP Calculus BC using zero LLMs — just a hand-encoded knowledge graph. This reframes the whole AGI-timeline debate. 🧵
5/ What's coming is hybrid: LLM for breadth, KG for precision. Both. The race to commodity single-modality ends this year.
6/ We're betting capital allocation on it.
getmanthan.com/briefing
Mayank Mathur retweeted
1/ Our weekly health check scans 3 fields for data-quality warnings. On 7 June it came back 0/0/0 — clean. It wasn't. Thread on the miss we found and haven't fully fixed yet.
1/ Our weekly health check scans 3 fields for data-quality warnings. On 7 June it came back 0/0/0 — clean. It wasn't. Thread on the miss we found and haven't fully fixed yet.
5/ Lesson: a self-healing loop is only as good as the agreement between what your writers emit and what your reader watches for. When those vocabularies drift quietly, the system doesn't fail loudly — it reports green.
6/ Fix isn't "widen the regex" blindly — some of those flags (insolvency proceedings) are legit business signals, not errors. Classify first, then converge the writers. Queued this week. Publishing the open defect now, not the tidy resolution later.
getmanthan.com
1/ April: a routine audit of our knowledge graph's sector labels found a 58% error rate on one batch. An LLM was matching keywords to sectors, not checking what companies actually did. We fixed it. Then the same bug came back somewhere else — read on.
6/ The claim: this bug resurfaces in every new field we classify from text, because it's structural, not a one-off. It did — 7 Jul, a home-improvement retailer almost got tagged "hardware" (as in electronics) because its own hardware meant bathroom fittings. Same guard caught it.
7/ If you run AI-assisted screening (fund, M&A, PE sector mapping, consulting sourcing tools) — ask your vendor for a keyword-match error they caught, not just their accuracy number.
getmanthan.com/live