pranav@harper ~ $ neofetch
██████╗ ███╗ ███╗ pranav @ san francisco
██╔══██╗████╗ ████║ ─────────────────────────────────────
██████╔╝██╔████╔██║ role forward deployed engineer, harper (yc w25)
██╔═══╝ ██║╚██╔╝██║ uptime 400+ prs · 30+ prod repos · since dec '25
██║ ██║ ╚═╝ ██║ kernel voice copilots, coding agents, event backbones
╚═╝ ╚═╝ ╚═╝ shell claude code, codex, cursor
prev salespatriot (yc w25) · sail · psu cs '25
Insurance has always run on agents. I build the new kind, at Harper (YC W25, $47M Series A, Series B underway). Mine are Dumbly, the realtime voice copilot our intake operators live in, and DaVinci, a background coding agent that drives work from research through integration testing, plus the pipelines and evals underneath both. The models were never the bottleneck. Getting an agent to behave when a carrier sends a scanned PDF sideways at 4:55pm on a Friday, that's the job.
- zero to one is where I'm useful. hand me a fuzzy problem and a customer who's annoyed, not a spec
- I build for someone's actual workflow, then sit next to them and watch where it breaks. the watching is the part people skip
- duct tape first, Temporal second. the trick is knowing which week you're in
| system | built with | what happened |
|---|---|---|
| Dumbly — realtime intake copilot | OpenAI Realtime API, Electron, packet-recommendation rules engine | 20+ operators, 300-400 intakes/day, manual intake down ~70% |
| DaVinci — background coding agent | ECS/Fargate, SQS FIFO, Step Functions, Postgres | research → plan → execute → integration-test; capacity-gated executors that root-cause failures (code vs test vs env) before retrying |
| lead lifecycle pipeline | Temporal (python + go workers), Postgres | raw lead → qualified opportunity → submission handoff. dual-write migrations, typed backfills, no downtime cutover |
| event backbone | schema registry, go SDK, NATS | event contracts for 5+ services. shipping a schema change through every consumer in one day |
| growth attribution stack | GCLID/msclkid capture, offline conversions, Customer.io | ad-click to closed-policy attribution; lifecycle campaigns fired off production events |
| reliability layer | otel, logfire, feature flags | fail-closed guards on anything customer-facing. about 50 PRs nobody sees, which is the point |
five open PRs and counting, all upstream of things I run in prod:
- anthropics/claude-agent-sdk-python #1087 — list-form system prompts crashed deep in the subprocess transport with a cryptic AttributeError, after the money was already spent. now fails fast at construction with an error that tells you what to do instead
- livekit/agents #6321 — a stored tool call with unparseable arguments could crash every later turn of a voice agent. fix + tests for the Anthropic/Google/AWS formatters
- inspect_ai #4418 — hidden
states were silently dropped from eval logs as
None. now they survive serialization - simonw/datasette #2826 — text and composite primary keys created through the JSON API were silently nullable, producing rows you can't view or delete. now NOT NULL at creation
- python-attrs/attrs #1584 — include/exclude filters matched attributes by equality, so excluding one class's field could silently drop an identical field on an unrelated class. now matched by identity
mine:
| project | one-liner |
|---|---|
| DaVinci case study | how the background coding agent works: the executor loop, the capacity state machine, and why the human gates are an architecture |
| autogen-distributed-agents | 20 concurrent agents on a gRPC runtime critiquing each other's ideas |
| langgraph-sidekick-assistant | worker/evaluator loop with memory, browsing, and a python REPL |
contributing next: openai-agents-python · pydantic-ai · semantic-conventions-genai
- Agents need a nervous system, not a bigger brain — the next gains come from durable execution and replayable state, not bigger models
- When both sides of the API are agents — B2B software when your customer's workflows are agentic too
- Human-in-the-loop is an architecture, not a checkbox — escalation ladders, fail-closed gates, and when the human is the product
| paper | notes |
|---|---|
| τ-bench | agents following domain policy with a human in the loop. closest thing to a benchmark of my day job |
| ReAct | still the skeleton under most agent loops, including mine |
| Reflexion | self-critique as memory. cheap and it works |
| CoALA | a map of agent architectures. I disagree with parts, which is why I keep rereading it |
| SWE-bench | made coding agents measurable |
| Generative Agents | the smallville paper. memory and routines for 25 agents in a toy town |
# toolchain.yaml
langs: [python, typescript, go, java, sql]
agents: [claude agent sdk + mcp, langgraph, autogen, openai realtime, evals]
voice: [openai realtime api, twilio, elevenlabs]
backbone: [temporal, nats, trigger.dev, n8n, rest + graphql]
data: [postgres + pgvector, supabase, qdrant, chroma, mongodb]
frontend: [react, react native, electron, sveltekit, next.js]
infra: [aws, gcp, cloudflare, docker, k8s, helm, argocd, github actions]
observability: [otel, logfire, posthog, grafana]
daily_drivers: [claude code, codex, cursor]most of my work lives in private org repos. the graph is what leaks through; the five PRs above are the part that doesn't have to.
$ curl -s pranavmishra.app | grep -i "building"