Backend & distributed-systems engineer · building RiskKernel (open-source agent reliability) and The Sunrize (AI trading co-pilot)
🔗 thesunrize.com · private beta
Built solo, end-to-end. Frontend, backend, agents, risk engine, broker integration, billing, infra.
| Component | What it does |
|---|---|
| Six-agent AI architecture | HELM (orchestrator) · PULSE (news) · PRISM (technicals) · HORIZON (macro) · SENTRY (risk) · LEDGER (auto-journal). Multi-model — Claude + GPT — with per-user spend caps. |
| Server-enforced risk engine | 2% per trade · 6% monthly cap · 15% position size · 3 concurrent positions. Hard-coded, no overrides. |
| Live broker execution | Zerodha Kite via official API. Smart bracket orders. Idempotent under network retries. |
| Nifty straddle scalper | Automated ATM straddle. Black-Scholes pricing, India-VIX-derived IV for paper mode. |
🔗 github.com/prashar32/riskkernel · Apache-2.0 · Go core + Python SDK
A self-hosted reliability runtime for AI agents. Deterministic cost / loop / time budgets and a kill switch — all enforced in compiled Go, never by an LLM.
| Capability | What it does |
|---|---|
| Deterministic enforcement | Cost, loop, and wall-time budgets enforced in compiled Go. No "the model decides when to stop." |
| Kill switch | Pull the plug at the runtime layer. Survives any prompt-injection or agent reasoning failure. |
| Crash-resumable runs | SQLite-backed run state. Restart anywhere in the loop. Git-native, owned memory. |
| Human approval gates | Side-effecting tool calls (writes, payments, deletes) require a human signature. |
| OpenTelemetry GenAI export | Plug into the OTel ecosystem; no proprietary lock-in. |
| One env var to adopt | BYO key, no telemetry, no signups. |
Adapters for Claude Agent SDK, OpenAI Agents SDK, and LangChain.
Positioning: the risk-engine / SRE layer for AI agents. Not a gateway (LiteLLM), not an observability dashboard (Langfuse), not a content-guardrail engine — interoperates with all of them.
"I've spent years building deterministic risk engines that wrap non-deterministic systems — the kind where a mistake costs real money. The safe part was never the smart part; it was the hard-coded layer around it. RiskKernel generalizes that discipline for AI agents."
First backend engineer when the team was 10. Stayed through it crossing 150.
|
Order Management Service 75K+ orders/day · 99.9999% uptime · first commit → launch Shipment Management Service — "the brain" 500K+ shipments/month · +35% delivery accuracy ClickHouse analytics layer 1M+ transactions/month · sub-second p95 Rule Management Service −50% feature rollout time · 800+ global sellers |
Centralized Logging Service Adopted by every backend team · cut platform-wide incident response 40% Monolith → microservices (+ DB migration) Led end-to-end · zero downtime, zero data loss Platform reliability SLO-driven monitoring, automated deploys, unified API versioning, 10× peak-traffic scaling IC scope Built platform tooling from concept to production as the team scaled 10 → 150+ |
Stack: Go · PostgreSQL · MongoDB · RocksDB · Redis · Kafka · ClickHouse · AWS
Most of my code lives in private org repos. Snapshot of recent activity →
|
|
|
| Platform | Status |
|---|---|
| Codeforces | Candidate Master · 2048 (top ~100 in India) |
| CodeChef | 6★ · 2208 |
| LeetCode | Guardian · 2249 (top 0.5%, all-time) |
| ICPC | 2× Regionalist |
| Google Code Jam | Round 2 qualifier (2021) |
| Meta Hacker Cup | Round 2 qualifier (2021) |
~3,000 problems solved. Set 400+ problems for InterviewBit / Scaler Academy hiring contests.
Independent evaluation work on frontier LLMs — rubric design, response grading, inter-rater reliability, eval pipelines.
👀 Open to founding-engineer · backend · AI-infra roles.
adarsh.prashar32@gmail.com · linkedin.com/in/adarshprashar · github.com/prashar32