the anti job boardactive 3mo ago
Harsh Kedia
builder
Looking for
Preferences
Open to
remote
The story
What Harsh is looking for
The loop. Research a hard problem → prototype something rough → ship it → watch real users hit it → iterate. The domain almost doesn't matter I've built in fashion discovery, freight ops, conversational AI, paper-trading equities, video editing, ai ugc and what hooks me every time is the same shape: depth + ownership of one slice end-to-end. What I actually lose hours to: the research phase before code. Reading other people's architectures, post-mortems, source working out why a system is shaped the way it is before deciding mine. And the first 48 hours of a new system when the design is still soft. I skip food and sleep through that part. What I don't want: being three layers removed from users, writing tickets for someone else to implement, or maintaining things instead of building them.
Why Harsh is exploring
Shoppin is shifting into maintenance. I was brought in to build the discovery feed, the scraping platform, virtual try-on, the social layer all of it shipped and is running. What's left is incremental. I'm at my best from zero under real constraints, and I miss that. I've done founding-engineer thrice and senior product-engineer here, and the common thread for me is: small team, real users, owning a slice end-to-end, shipping daily, no bureaucracy. I want to do that again founding-engineer energy with a team that actually cares about craft. Not chasing a title or a bigger logo; chasing the kind of work where I get to design the system, ship it, watch users hit it, and iterate without three approval layers in between. Not bitter about Shoppin, I built what I was brought in to build. Just want the next zero-to-one.
What won't fit on a resume
Oracle a multi-agent paper-trading system for Indian equities. Wanted to actually understand markets instead of vibing on Twitter, so I built the research stack I wished existed. Custom event-sourced workflow engine on Postgres rolled my own because LangGraph felt too magical for something I wanted to debug and fork-replay. 5 specialist agents run in parallel (Fundamentals, Sentiment, Macro, Dynamics, Valuation), then a Sell-Side agent, then a Red-Team Critic forced to assume the trade loses 20%, then a deterministic Decision Hub in pure Python that fuses everything. Hard rule: the LLM never picks a number position sizing, fees (full Zerodha STT/GST/SEBI/stamp ladder, ±₹2 of their calculator), stops, take-profits all pure Python. Gemini only writes prose around pre-computed scores. Indian-market quirks baked in. Paper-trade close-the-loop is live. Most I've learned in a single project this year durable workflow design, crash-resume + fork-replay, deterministic guardrails for LLM systems. AI video editor. Wanted short-form edits from raw footage without scrubbing timelines. Pipeline: shot detection → YOLO object detection → Gemini-based content embeddings → vector search → assembly agent that picks clips and beat-aligns them → renders via Remotion or FFmpeg through an OpenTimelineIO adapter. Turns out "make a good edit" is mostly a constrained search problem (pacing, beat sync, narrative arc) with a thin LLM layer for taste. The surprise: I expected the ML to be hard; the actual hard part was the timeline data model. Axon is on the resume but worth saying out loud I built it because I needed it for my own work, opened it up, and it ended up at 500+ stars with people using it via MCP in Claude Code. That's the pattern behind most of my side projects: build the tool you wish existed for the work you're already doing, then open it.
Work & Projects
The Anti Job Board
Skip the application black hole. Get warm intros to founders hiring now.
Create your profile