I don't just use abstractions.
Sometimes I build what's underneath them.
┌─────────────────────────────────────────────────────────┐
│ │
│ ANSH SURANA │
│ │
│ ▸ AI systems │
│ ▸ Retrieval & ranking │
│ ▸ Backend infrastructure │
│ ▸ Security engineering │
│ ▸ Edge / Android AI │
│ ▸ Developer tooling │
│ │
│ Current loop: │
│ │
│ BUILD → MEASURE → BREAK → FIX → SHIP │
│ │
└─────────────────────────────────────────────────────────┘
I like working on the layer where things stop being magic and start becoming engineering.
Vector indexes. Retrieval pipelines. Authentication protocols. Agent execution. Android internals. Real-time systems.
Basically:
if something interesting is happening underneath the API, I probably want to know how it works.
vector-db-from-scratch
RAGINGOA
Bitfrost
Jagrut
Sentinel
Job Recruitment
HNSW · WAL · Snapshots · Recovery · PQ · Filtering · Concurrency
This is probably the project that best represents how I like to build.
Instead of wrapping an existing vector database, I implemented the machinery underneath it.
┌───────────────┐
│ Client │
└───────┬───────┘
│
▼
┌───────────────┐
│ API Layer │
└───────┬───────┘
│
┌──────────┴──────────┐
▼ ▼
┌──────────────┐ ┌──────────────┐
│ HNSW Search │ │ Filtering │
└──────┬───────┘ └──────┬───────┘
│ │
└──────────┬──────────┘
▼
┌───────────────┐
│ Vector Store │
└───────┬───────┘
│
┌─────────┴─────────┐
▼ ▼
┌────────┐ ┌──────────┐
│ WAL │ │ Snapshot │
└────┬───┘ └─────┬────┘
└─────────┬──────────┘
▼
Crash Recovery
- HNSW approximate nearest-neighbor search
- optimized distance calculations
- metadata filtering
- tombstones and compaction
- product quantization
- WAL persistence
- snapshot recovery
- concurrent read/write workloads
- actual recall and latency gates
- crash-recovery testing
Rust HNSW WAL PQ Concurrency
FAISS HNSW · Embeddings · Reranking · Guardrails · Multilingual · STT
The pipeline is designed around measurable retrieval latency and layered safety rather than a single retrieve → prompt → answer call.
Input
│
▼
Language / STT
│
▼
Guardrails
│
▼
Embedding
│
▼
FAISS HNSW
│
▼
Hybrid Retrieval
│
▼
Reranking
│
▼
Evidence
│
▼
Generation
│
▼
Post-check
│
▼
Response
The repository includes reproducible retrieval benchmark artifacts with measured latency breakdowns and percentile statistics.
Python FAISS HNSW Embeddings Reranking
HMAC · Replay Protection · Rate Limiting · Circuit Breaker · Cache · Telemetry
The interesting part isn't forwarding an API request.
It's everything that can go wrong around one.
Client
│
▼
Identity
│
├── HMAC
├── Timestamp
├── Replay Protection
└── Tenant Validation
│
▼
Policy Layer
│
┌────┴────┐
▼ ▼
Cache Upstream
│ │
└────┬────┘
▼
Telemetry
Go HMAC-SHA256 Supabase WebSockets Gateway
Wake Word · ONNX/TFLite · Hinglish · Accessibility · Camera2 · Automation
Instead of stopping at conversational AI, Jagrut connects language understanding to real Android capabilities.
Voice
│
▼
Wake Word
│
▼
STT
│
▼
Command Parser
│
├───────────────┐
▼ ▼
Local Action AI Planner
│ │
└───────┬───────┘
▼
Action Validation
│
▼
Android Execution
Built around Android internals such as AccessibilityService, Camera2, and on-device model runtimes.
Kotlin Android ONNX Runtime AccessibilityService Camera2
Sentinel combines academic document processing with a structured claim relationship model.
The interesting bit isn't simply generating an answer.
It's being able to reason about:
Where did this claim come from? What supports it? What depends on it? What is connected to it?
Source
│
▼
Document
│
▼
Claim
│
┌┴─────────────┐
▼ ▼
Parent Children
│ │
└──────┬───────┘
▼
Evidence Graph
Next.js React Prisma Supabase Gemini
Rather than asking an LLM to simply “score this candidate,” the system separates retrieval and ranking into distinct stages.
Candidate
│
├───────────────┐
▼ ▼
Lexical Semantic
Retrieval Retrieval
│ │
└───────┬───────┘
▼
Hybrid Score
│
▼
Deterministic Rules
│
▼
Cross-Encoder
│
▼
Final Ranking
Python BM25 Embeddings Cross-Encoder Docker
Not every experiment deserves a giant section.
But they're here:
| Project | What it explores |
|---|---|
| Collaborative Workspace | Real-time collaboration, synchronization, whiteboards & shared state |
| DevLens | AI-assisted repository analysis and developer tooling |
| ChronoGuard | Computer vision, tracking, privacy filtering & real-time events |
| Sark Pharma Tech Services | Deployed full-stack product |
┌────────────────────────────────────────────────────────┐
│ │
│ BUILD │
│ ↓ │
│ MEASURE │
│ ↓ │
│ FIND THE UGLY EDGE CASE │
│ ↓ │
│ BREAK IT │
│ ↓ │
│ FIX IT │
│ ↓ │
│ SHIP │
│ ↺ │
│ │
└────────────────────────────────────────────────────────┘
I genuinely enjoy the part where the happy-path demo stops working.
That's usually where the interesting engineering starts.
