ABHISHEK ACHARYA — SYSTEM BOOT
Skip to content
A. ACHARYA

Kathmandu, Bāgmatī, NepalAVAILABLE FOR WORK

AI & MLENGINEERAUTOMATION

I build retrieval-augmented systems, wire language models into real infrastructure, and automate the workflows around them.

AI CORE
INITIALISED
RETRIEVAL
INDEXED
AUTOMATION
READY
SEE THE WORKGET IN TOUCH
01ABOUT

I build systems that
retrieve, reason & automate.

I work across AI/ML, natural language processing and full-stack engineering — most often on systems that need to retrieve the right context before they answer.

At Nepal Telecom I engineered a bilingual customer-support assistant handling data-package and service enquiries in Nepali and English, using FAISS and sentence-transformers for retrieval and a live CMS API for real-time package data.

At FlickerArts I led development of FilmFinance, a SaaS movie finance tracker — budgeting, cash flow, vendor procurement and approvals — built on Node.js, PostgreSQL and TypeScript.

The through-line is the same: retrieval, orchestration, and the backend architecture that keeps both honest in production.

02AI / MACHINE LEARNING

Retrieval first.
Then generation.

A language model that answers from memory will invent policy, prices and facts. Every system here grounds its answers in retrieved context — indexed, versioned and rebuildable.

  1. 01

    DATA

    Scraped corpora, documents, CMS APIs

    OK
  2. 02

    CLEANING

    Parsing, normalisation, deduplication

    OK
  3. 03

    CHUNKING

    Document segmentation for retrieval

    OK
  4. 04

    EMBEDDING

    all-MiniLM-L6-v2 sentence-transformers

    OK
  5. 05

    INDEXING

    FAISS IndexFlatL2 · ChromaDB · Qdrant

    OK
  6. 06

    RETRIEVAL

    Semantic search over vector store

    OK
  7. 07

    GENERATION

    Gemini / Ollama grounded in context

    OK
  8. 08

    SERVING

    Django REST · Express · WebSockets

    OK

STACK IN USE

  • Python
  • FAISS
  • ChromaDB
  • Qdrant
  • Sentence-Transformers
  • Google Gemini
  • Ollama
  • Cohere
  • LangChain
  • NumPy
  • Pandas

WHY THREE VECTOR STORES

FAISS for a local, file-backed index that ships with the app. ChromaDB when persistence and metadata filtering matter more than raw speed. Qdrant when retrieval has to survive as its own service behind a queue.

STATIC VS. LIVE

Indexed documents answer what does not change. Anything price- or status-sensitive is a tool call at request time — the model never quotes a number it read during ingestion.

03AUTOMATION

The work that
runs itself.

Scheduled aggregators, document pipelines, queued workers and event-driven workflows — the layer that turns a model into something that operates without anyone watching it.

BATCH AND POLL

Long scrapes cannot block an execution. Submit the batch, wait, poll the status, branch on completion — the workflow yields instead of holding a connection open for minutes.

QUEUED WORKERS

Celery with Redis for scheduled Git operations; background workers for document ingestion. Ingestion returns immediately — chunking and embedding happen out of band, so one large upload never starves the API.

DETERMINISTIC BY DEFAULT

Extraction that has to be auditable does not use a model. Invoice and purchase-order fields come out of a maintained regex library, and candidate scoring sits outside the LLM entirely — same input, same output, every run.

EVIDENCE OVER CONFIDENCE

A score carries the text that produced it. A 0–100 tender relevance rating or a candidate match breakdown shows the matched terms, so a human can overrule it for a stated reason rather than a hunch.

WORKFLOW NODES

  • Schedule TriggerNODE

    Daily and weekly cycles across tender sources

  • HTTP / RSSNODE

    ADB, World Bank, AIIB and SAM.gov feeds and APIs

  • WebhookNODE

    Entry point — synchronous response path

  • Code (JS)NODE

    Regex extraction, parsing, merge, relevance scoring

  • Split In BatchesNODE

    Bounded chunks so token spend stays predictable

  • AI AgentNODE

    Claude and Gemini structured extraction

  • Data TablesNODE

    State that survives between runs — PostgreSQL-ready

  • SharePoint / OutlookNODE

    Enterprise file writeback and digest delivery

  1. 01

    TRIGGER

    Schedule · webhook · Celery Beat

    OK
  2. 02

    COLLECT

    RSS · REST APIs · SharePoint · Firecrawl

    OK
  3. 03

    EXTRACT

    Regex library · PDF/DOCX text · OCR fallback

    OK
  4. 04

    NORMALISE

    Parse, merge, dedupe, batch

    OK
  5. 05

    SCORE

    Evidence-based relevance · deterministic matching

    OK
  6. 06

    ACT

    Data tables · Outlook digests · SharePoint writeback

    OK
04SELECTED WORK

Things that
actually run.

Each case study describes the real architecture — every component listed exists in the repository or the deployed workflow. Where numbers appear they are measured in operation, never invented.

ALSO BUILT — PRIVATE REPOSITORIES

  • MultiAgentAISystem

    Multi-agent research system — planner, researcher, retriever and summarizer agents over a RAG layer with web-search and scraper tools.

  • PasalHub

    Multi-tenant SaaS webstore platform for Nepal — Django REST, PostgreSQL, Node/Socket.IO real-time service, host-resolved merchant storefronts.

  • Auto Git Scheduler

    Scheduled Git automation — Django, Celery, Celery Beat, Redis and GitPython, containerised with Docker Compose.

SOURCE AVAILABLE ON REQUEST

05STRATEGY

Think three
moves ahead.

Building intelligent systems is like playing chess. Every decision changes the state of the system, and every action creates the next set of possible moves.

STRATEGYDECISIONLEARNINGAUTOMATION

06EXPERIENCE

Where the work
has happened.

  1. June 2026 — Present3 monthsCURRENT

    AI & Automation Engineer

    K&A Power

    • Built a multi-source tender radar in n8n, aggregating ADB, World Bank and AIIB consulting notices on a daily schedule with bilingual English/French relevance scoring.
    • Specified an AI-powered recruitment platform across thirteen modular workflows over a normalised PostgreSQL schema, using Claude for structured CV extraction and a deterministic scoring layer outside the model.
    • Built an invoice and purchase-order reconciliation pipeline — regex-based extraction from SharePoint PDFs, multi-key linkage, variance detection and metadata writeback.
    • n8n
    • PostgreSQL
    • Claude API
    • JavaScript
    • SharePoint
    • Microsoft 365
  2. March 2026 — April 20262 monthsKathmandu

    Project Lead Developer

    FlickerArts

    • Led development of FilmFinance, a comprehensive movie finance tracker SaaS web application, enhancing project management capabilities.
    • Managed multiple modules including Dashboard, Team Management and Finance, ensuring seamless integration and user experience.
    • Streamlined financial processes such as budgeting, cash flow management and vendor procurement, optimizing operational efficiency.
    • Node.js
    • PostgreSQL
    • TypeScript
  3. March 2026 — April 20262 monthsKathmandu

    Project Manager

    FlickerArts

    • Designed a SaaS-based movie finance tracker utilizing Node.js, PostgreSQL and TypeScript.
    • Collaborated with cross-functional teams to ensure seamless integration and functionality.
    • Focused on enhancing finance tracking for movies and similar projects, improving project management efficiency.
    • Node.js
    • PostgreSQL
    • TypeScript
  4. January 2026 — March 20263 monthsKathmandu

    AI/ML & NLP

    Nepal Telecom

    • Engineered an AI chatbot handling customer inquiries about data packages and services in both Nepali and English.
    • Leveraged FAISS and sentence transformers to ensure users received accurate and timely information.
    • Integrated with NTC's CMS API for real-time updates, significantly enhancing user experience.
    • Python
    • FAISS
    • Sentence-Transformers
    • Django
    • Ollama
07SKILLS

A connected
stack.

Hover any technology to see what it connects to. Everything listed is evidenced by a repository or a role — nothing aspirational.

49 TECHNOLOGIES

08CONTACT

Let’s build
something
intelligent.

Open to AI/ML and automation engineering work. The fastest way to reach me is email.

javascript1415@gmail.com