╔═══════════════════════════════════════════════════════════════════════╗
║ ____ _ ║
║ | _ \ __ _ _ __ __ _ _ __ ___ ___ ___ __| | ___ ___ ║
║ | |_) / _` | '__/ _` | '_ ` _ \ / __/ _ \ / _` |/ _ \/ __| ║
║ | __/ (_| | | | (_| | | | | | | (_| (_) | (_| | __/\__ \ ║
║ |_| \__,_|_| \__,_|_| |_| |_|\___\___/ \__,_|\___||___/ ║
║ ║
║ Autonomous Agents • Terminal TUIs • RAG • Systems Engineering ║
╚═══════════════════════════════════════════════════════════════════════╝
$ cat ~/.profile.json
{
"engineer": "Paramveer Singh",
"focus": ["Autonomous Coding Agents", "Hierarchical RAG", "Distributed Systems & TUIs"],
"philosophy": "Understand the runtime from scratch rather than wrapping black-box SDKs"
}$ ls -la ~/projects/featured --sort=recent| System / Tool | Core Architecture & Problem Solved | Stack |
|---|---|---|
| Maggie | Autonomous terminal coding agent. Dual PLAN/BUILD modes, multi-provider LLM gateway, human-in-the-loop security confirmation gates, reactive @ file mentions, and persistent SQLite WAL sessions. |
Bun • OpenTUI • SQLite |
| Nexus | Multi-agent PR reviewer. Native GitHub Action running AST diff parsing → planner → 4 parallel specialist reviewers (style, security, perf, tests) → critic agent → aggregator with LLM-as-judge evals. | TS • Bun • Groq |
| ATLAS | Production-grade hierarchical RAG. Docling ingestion, Jina v3 embeddings, RAPTOR tree clustering (UMAP + HDBSCAN), hybrid BM25 + pgvector HNSW search with RRF fusion, and Prometheus observability. | Bun • Fastify • pgvector • Prometheus |
| Night Code | Open-source terminal coding harness. Local code inspection, AST edits, and token-by-token streaming via a stateless relay with local execution authority. | TypeScript • OpenTUI |
| BackendSaga | Interactive systems failure explorer. Incident triage engine, failure modes knowledge graph, and evidence library for distributed backend systems. | Next.js • TypeScript • GraphViz |
| Ben | Systems agent runtime in Rust. Learning-oriented terminal coding agent harness exploring low-level process management and tool orchestration. | Rust |
| Junior Doctor | AI clinical document pipeline. Scanned patient records → structured discharge summaries with provenance tracking, conflict detection, and anti-hallucination guardrails. | Python • LLM Evals |
| devops-infra | Distributed inference architecture. Multi-VM inference cluster on AWS with private subnet routing, RPC worker communication, and Terraform provisioning. | Python • Terraform • AWS |
🪄 Maggie — Autonomous TUI Coding Assistant
- Dual Execution Modes: Instant
Shift+Tabtoggle between PLAN (read-only architectural analysis & codebase reconnaissance) and BUILD (autonomous file modifications and command execution). - Security Confirmation Gate: Destructive actions (
write,edit,bash) trigger human-in-the-loop confirmation gates before execution. - Universal Clipboard & Reactive Mentions: Native OSC 52 mouse drag text selection copies directly to system clipboard across SSH/local sessions; reactive
@file search with fuzzy scoring. - Persistence: Bun SQLite in WAL mode with cascading session indexing and cross-session resume.
🛡️ Nexus — Multi-Agent PR Reviewer Action
- Specialist Fan-Out: Diff parser slices changes by AST boundary and fans them out to 4 parallel specialist reviewer agents (Style, Security, Performance, Test Coverage).
- Writer / Critic Loop: A dedicated Critic agent validates each finding against the original file context before aggregation to eliminate hallucinated issues.
- LLM-as-Judge Evals: Benchmarked against synthesized PR datasets with ground truth validation to track precision/recall over prompt and model iterations.
🧠 ATLAS — Production-Grade Hierarchical RAG Agent
- RAPTOR Hierarchical Indexing: Replaces flat chunking by recursively clustering document vectors (UMAP dimensionality reduction + HDBSCAN) and summarizing clusters with Gemini to answer high-level thematic queries.
-
Hybrid Retrieval & RRF: Combines pgvector cosine similarity (HNSW index) and PostgreSQL BM25 full-text search using Reciprocal Rank Fusion (
$k=60$ ). - Precision Reranking: Passes candidate documents through Jina Reranker v2 before final prompt synthesis.
-
Observability: Prometheus
/metricsendpoint with Grafana dashboards for latency, token spend, and retrieval cache hits.
[Core Languages] TypeScript • Rust • Python • C++ • Go • SQL
[Agents & Retrieval] OpenTUI • Jina v3 • pgvector • RAPTOR • Docling • BM25
[Backend Runtimes] Bun • Node.js • Fastify • FastAPI • Express
[Databases & Caching] PostgreSQL • Redis • SQLite (WAL) • Neo4j
[Infra & DevOps] Linux • Docker • Terraform • AWS (VPC/Private Subnets) • GitHub Actions
