Mittal — How a Private Chef Startup Went All In on AI Agents (YC Root Access)
Yhangry is a private chef marketplace doing $15 million in GMV that’s going all in on AI agents across every function of the company — from an autonomous bug fixer that shipped 25 fixes in its first week to an AI product that matches chefs and customers instantly.
In this recent batch talk, founder Siddhi Mittal walks through three real business use cases for agents at yhangry, how she turned teaching AI in plain English into a growth channel worth $50K in free conference slots, and the hard org decisions she made to rebuild the company as AI native from the ground up.
— channel description, YC Root Access
TL;DR
A ~4:55 micro-talk (one of several short founder-cards in a recent YC batch session on the YC Root Access channel; published 2026-05-19, same day as the Garg / AnswerThis talk on the same channel). Siddhi Mittal is the founder of Yhangry (“WhyHungry” in the ASR), a private-chef marketplace. The channel description sets the operating scale at $15M GMV; in the talk Mittal puts it at “roughly $50 million GMV trying to like 10× this year” — likely indicating GMV growth between the description draft and recording, or possibly an ASR artifact.
The substantive contributions are five:
1. Autonomous bug fixer (~0:14–1:01). “We had a backlog of bugs… what eventually gets dropped are tiny little bugs, the really annoying ones that our developers just never get to.” Built an autonomous bug fixer in less than 4 days; 25+ bugs fixed in week one and shipped. One-shot bug-fix pass rate: “60 to 70% right now.” The learning: “how we can feed it enough context such that it’s self-improving.” This is the wiki’s smallest worked example of a self-extending-agent-harness — Mittal does not name a thin-harness / fat-skills architecture explicitly, but the self-improving-context framing is the same architectural concern Garg’s source articulates with the agent-editable instructions.md mechanism.
2. Founder-brand-as-growth via teaching AI agents in plain English (~1:01–2:08). “Talking and teaching about AI agents in plain English. I realized that’s the edge, because no one has a clue as to what’s going on in this whole world.” Pitched “instead of pitching Yhangry, I’m going to teach everyone how to build AI agents in 30 minutes” to conference organisers. Got $50K worth of conference slots for free. The teaching deck embeds an affiliate-link integration and a demo of the new Yhangry AI product. “Screenshot the slide, chuck it into Claude Code, learn about it later” — the audience then posts on LinkedIn about it. “It’s like win, win, win, win, win.” The wiki’s first founder-vantage articulation of distribution-by-teaching-the-underlying-technology, where the founder’s domain expertise becomes the growth channel.
3. Yhangry AI product — instant chef-customer match (~2:08–3:19). The current Yhangry product “feels very old school — finding a chef, booking, so many steps and so much back-and-forth, we are losing a lot of people.” Mittal’s framing of the AI product: “It’s like Claude for chefs. We have all the data — what chefs respond to, what customers like — we just never knew how to match them instantaneously and make sure the one-shot match is really good.” Status: validated on chefs (“everyone wants it”); the launch blocker is divergence on the chef side — “some chefs are very by the book, some chefs are lost in the matrix.” MVP design is in-flight.
4. The fire-and-rehire-in-a-week org move (~3:25–4:00). Mittal returned from maternity leave (“I had a baby 3 months ago, I got hooked on OpenClaw, I was using voice Telegram to the point where I nearly separated from my husband — we’re okay now”) and immediately rebuilt the engineering team. “I fired my tech lead because I realized he did not know what skills was, and he was the ceiling in our company. And I re-hired our new head of engineering all within a week. In March, we are really all in.” The bluntest articulation in the wiki corpus of the AI-native-rebuild’s organisational cost — “I probably need to kill my human empathy a little bit more, a little bit faster.”
5. Weekly agentic labs as the internal-onboarding ritual (~4:09–4:31). “Everyone has a different learning curve right now and it’s really kind of wild. Like people are saying they’re building agents, but you just don’t know how good they are. So we’re trying to do this — me and the head of engineering together. Like literally drawing out stuff and getting Claude code to just use all the transcripts and chuck it back in nice diagrams everyone understands after.” This is the wiki’s first articulation of weekly cross-team agentic learning sessions as a deliberate internal-onboarding mechanism — paralleling Anthropic 2026’s account of agentic-coding-onboarding at Anthropic, but at small-startup scale.
Founder backstory. Mittal’s domain knowledge: Columbia AI degree from 2013, then operating Yhangry. “My domain knowledge is translating my AI Columbia degree from 2013 and ability to grasp concepts in plain English.” The 13-year-gap-from-Columbia-AI to now-it-actually-matters is itself worth noting as a durable-skills datapoint — the plain-English explanation skill compounded across non-AI roles, then became the growth channel when the rest of the world needed translation.
Caveats. YC marketing channel; ~5 minutes total; one of multiple short founder-cards in a batch session (the description’s “recent batch talk” implies several speakers in the same event). Numerical claims (25+ bug fixes / 60–70% pass rate / $50k of free conference slots / $15M-or-$50M GMV) are unaudited founder self-reports. The fire-and-rehire-in-a-week claim is presented as a triumphant org-redesign vignette but tells us very little about how Mittal handled the labour-law / employee-experience side of that decision — read with caution.
Why this matters in the corpus
This talk is the most candid founder-vantage articulation in the wiki on the org-pain side of the AI-native rebuild. The “I fired my tech lead… I probably need to kill my human empathy a little bit more, a little bit faster” lines complete a corpus pair with CS153’s smooth “yeah there are people who are still operating like co-pilot level from last year and it’s like not going to make it, bro.” Same observation — operators who don’t know the current tooling are the ceiling — but Mittal’s version is what actually doing the firing sounds like, told live from inside the rebuild.
The founder-brand-as-distribution-by-teaching angle is the wiki’s first articulation of a specific 2026 growth pattern: when a founder can credibly translate the AI tooling into plain language, the act-of-teaching itself becomes the customer-acquisition surface — better than pitching the company, because the teaching gives the audience real value and the company demo is embedded in the teaching deck. This is a structural counter to the “AI-native means cold automation” default framing — Mittal’s growth channel is a deeply human, in-person teaching gesture that the AI-native rebuild made possible.
What was actually ingested
The full ~4:55 transcript was read end-to-end; the talk is compact enough that no chapter-level filtering was needed. The transcript renders Yhangry consistently as “WhyHungry” (ASR phonetic interpretation), and a “Tom” reference (~3:58 — “This is like the oral question Tom was asking”) is likely Tom Blomfield (YC partner per the Y Combinator entity page’s open-question list of speakers in YC Root Access multi-talk sessions). The talk appears to be one slot in a multi-speaker YC batch event ingest of which Garg’s talk is another.
Linked entities and concepts
Entities promoted by this source:
- Y Combinator — channel; already entity. Bumps source-count.
Dangling — single-source mention, deferred:
- Yhangry — private-chef marketplace; the talk doesn’t state a YC batch year. First wiki mention; promote on second source.
- Siddhi Mittal — founder of Yhangry; Columbia AI degree (2013). First wiki mention.
- Tom (Blomfield?) — YC partner referenced in passing (“the oral question Tom was asking”); likely the same Tom Blomfield already on the Y Combinator entity page’s open-question list. The Mittal mention is the second-source mention by surname-less Tom-as-a-YC-partner-name but doesn’t itself substantiate the surname; the Y Combinator entity-page Tom-Blomfield slot stays Dangling pending an unambiguous second source.
Concept pages touched:
- agentic-engineering — adds the autonomous-bug-fixer with 60-70% one-shot pass rate, self-improving via context as a small-startup-scale worked example.
- ai-employment-effects — adds the fire-and-rehire-in-a-week org-redesign vignette as the wiki’s bluntest articulation of AI-tool-fluency-as-firing-criterion at the technical-leadership layer.
- durable-skills — adds the plain-English-explanation-as-distribution worked example at founder scale; 13-year-gap-from-Columbia-AI as a long-arc compounding of a non-AI-coded skill that became AI-coded after the rest of the world caught up.
Source quality
- Channel: YC Root Access — recent-batch-talk channel.
- Format: ~5-minute slide-driven solo founder talk in a multi-speaker session; one of multiple ingestible founder cards in this YC Root Access event.
- Empirical anchors: $15M GMV (channel description) / $50M GMV (Mittal’s voice line — discrepancy unresolved); 25+ bug fixes in week one; 60–70% one-shot pass rate; $50k worth of conference slots; 4-day autonomous-bug-fixer build time. All founder self-reports; small scale enough that nothing is hugely consequential but the fire-and-rehire-in-a-week claim is the most narratively load-bearing.
- Bias / motive: Founder-marketing in a YC batch session; the talk’s “I’m in the middle of this figuring it out” posture buys some authenticity over the more polished YC-CEO-anchor lectures, but everything is still on a YC-channel.
- Transcript provenance: yt-dlp VTT fallback; ASR rolling-caption duplicates de-duplicated by the build script’s suffix-overlap collapse pass; Yhangry consistently mis-spelled as “WhyHungry” in the raw — corrected against the channel description in attribution.