Ashray Adhikari

sqlite-hybrid-search

GitHub PyPI

In-process C++17 library that fuses dense HNSW vector search (usearch) and BM25 full-text retrieval (SQLite FTS5) via Reciprocal Rank Fusion with per-chunk exponential recency decay. Single-writer/concurrent-reader WAL concurrency, no external locking. Pimpl public API. nanobind Python bindings with prebuilt wheels for Linux x86-64 and macOS Apple Silicon.

Cold-start time cut from 14.6 seconds to ~40ms by persisting the HNSW graph to disk. The whole retrieval stack fits in-process — no network hop, no separate service.

sub-0.2ms dense p50 latency
at 100k documents

moat

GitHub

SEC 10-K/10-Q RAG application. In production, with Stripe billing and real users. Not a demo.

pgvector dense retrieval + Postgres FTS via RRF. Cohere/CrossEncoder reranking. LangGraph agent with XBRL financial fact tools. HITL review gates. Langfuse observability throughout.

moat-blond.vercel.app ↗

Python · FastAPI · LangGraph · Next.js · Supabase/pgvector · LiteLLM · Stripe · Langfuse

LangGraph for the agent loop — explicit state graph rather than framework magic, every branching decision legible and testable

RRF over pure dense search because financial filings have both precise entity names (dense-friendly) and long analytical passages (BM25-friendly)

HITL gates at the reranking stage because hallucinated financial figures have real consequences

NFT Munich

nftmunch.club ↗ GitHub

Full-stack web presence for NFT Munich football club — a real sports club with fixtures, squads, and club management needs. Sole developer across design, implementation, and production deployment.

CMS-backed so the club can update fixtures, news, and squad pages without touching code. Live in production at nftmunch.club.

Next.js · TypeScript · Tailwind CSS · Headless CMS · Vercel

Headless CMS so non-technical club staff can publish match results and news independently — the admin interface was part of the deliverable

Sole developer on a real client project: design direction, component system, CMS schema, and deployment all owned end-to-end

Static generation for content pages, server rendering for dynamic fixtures — keeps the club's page fast without a backend team to maintain it

mlscan

GitHub

Static ML supply-chain scanner. Simulates the pickle VM stack and memo to recover imports hidden through multi-step opcode chains — without executing any input. Zero external dependencies.

Formats

safetensors · GGUF · HDF5 · ONNX · Keras

Detects

Jinja2 sandbox escapes · Zip Slip · shadow weights

Typosquat engine

NFKC + homoglyph map + casefolding

Output

SARIF · CycloneDX ML-BOM

Performance

Parallel 8-thread scanning

Dependencies

None. Python stdlib only.

ragaudit

GitHub

Three-tier adversarial evaluation platform for LLM and RAG systems. Next.js frontend, Python FastAPI backend, Go worker orchestrating Docker sandboxes for isolated tool execution. Maps findings to OWASP LLM Top 10 2025.

Per-tool HITL approval gates with three decisions — allow, deny, or allow-with-network-access. Strict sandbox mode fails closed if the worker is unavailable; enterprise profile disallows auto-approve entirely. Each finding captures the complete exchange: prompts, multi-turn history, tool calls, raw outputs.

Execution modes

Internal agent · external HTTP API · mock provider

Model providers

Ollama · OpenAI · Anthropic · Groq · Azure · Gemini

Export

JSON · SARIF · PDF · HTML

Coverage

OWASP LLM Top 10 2025

Stack

Next.js · TypeScript · Python 3.12 · FastAPI · Go · Docker · PostgreSQL · Redis

Secrets

Vault key_ref abstraction — never logged or transmitted in plaintext

Atlas-Workbench

GitHub

Multi-tenant AI workspace platform. Five specialized agents behind a single interface, plus a sandboxed developer environment with a real terminal and git integration — isolated per-user containers, not a shared shell.

Research agent

Web search · source synthesis · citation tracking

Doc Q&A agent

Uploaded document retrieval with inline source citations

SQL agent

Natural-language to SQL with schema introspection

Code review agent

Diff-aware analysis, security checks, style feedback

Dev assistant

Terminal access + git in per-user sandboxed container

Infrastructure

Docker Compose · Kubernetes · multi-tenant isolation

dossier-agent

GitHub

01async Celery ingestion

02AstraDB dense + BM25 hybrid retrieval

03dynamic routing (docs ↔ web search)

04RAGAS evaluation

05SSE streaming response

5-stage LangGraph pipeline that routes dynamically across uploaded documents and live web search depending on query type. AstraDB provides dense vector retrieval; BM25 covers exact-match queries that embeddings tend to miss.

RAGAS evaluation pipeline runs continuously — faithfulness, answer relevance, and context precision tracked per query. SSE streaming so responses feel immediate even on long retrievals.

injection-guard

GitHub

01normalization fan-out

0217-signal heuristic family

03MiniLM-L6 embedding-margin

04DeBERTa-v3-small + Platt cal.

05output & canary checks

ROC-AUC 0.9978
PR-AUC 0.9965
97.2% recall @ 1% FPR

Prompt-injection detection cascade. Each tier is an independent classifier — the cascade exits early on high-confidence decisions. 88% cache hit rate compresses average latency to 2.73ms/chunk (p95 46.9ms).

Trained on 2,100 examples spanning 9 attack classes across 11+ languages — payload splitting, indirect injection, obfuscation. Group-split evaluation prevents leakage between attack variants. Test set n=848.

llm-serving-from-scratch

GitHub

LLM inference server study on Qwen2.5-1.5B-Instruct running on Apple Silicon MPS. Four serving architectures implemented in sequence to isolate the contribution of each decision. Fused PagedAttention kernels deliberately excluded so each architectural effect is independently measurable.

naive single-lock → static batching → continuous batching → paged KV + prefix cache

+39% throughput (10.86 → 15.10 tok/s) −36% p50 latency (93.6 → 60.0s)

llm-redteam-harness

GitHub

Adversarial testing framework for LLMs. 13 techniques × 6 payloads factorial design, 30 trials per cell, 2,700 total attempts against llama3.2:3b. Three-judge ensemble: exact canary match (κ=1.00), LLM judge (κ=0.62), regex (κ=0.38). Wilson score intervals throughout — per-payload analysis, no pooling that would mask payload-specific effects. Benign control arm catches refuse-everything models.

Technique	ASR	Finding
format_smuggle	100%	Dropped to 43% with benign preamble — phrasing substantially affects measured ASR
roleplay_dual_channel	90%
fiction_scene	51%
authority_developer	50%
structured output (JSON schema)	100%	Bypassed policy entirely — discovered post-study, not in original 13

AshrayAdhikari

Work

Projects

Ashray Adhikari

Ashray
Adhikari