e
e_parashumti

Erjon P

@e_parashumti

Software Developer

Kosovo
Inglês, Albanês
Algumas informações são exibidas no idioma inglês.
Sobre mim
I am a software developer proficient in Python, Java, and TypeScript with a keen focus on AI development. As a Master’s student in Machine Learning and Data Science, I have hands-on experience automating workflows, building prototypes, and collaborating to solve real-world problems. I am passionate about creating innovative, user-centered solutions in dynamic environments.... Saiba mais

Habilidades

e
e_parashumti
Erjon P
offline • 

Conheça meus serviços

Integrações de IA
I will build ai chatbots, agents, and llm integrations
A partir deUS$ 100 /projetoMais detalhes

Experiência profissional

ISA

Software Developer

ISA • Período integral

Mar 2026 - Present • 7 mos

Software developer at Reveal, an enterprise e-discovery platform, maintaining connectors that collect content and metadata from business SaaS tools. - Maintain data-collection connectors for Slack, SharePoint, OneDrive, Salesforce, Okta, Confluence, Office 365/Exchange, and Zoom, enabling reliable ingestion for downstream e-discovery - Added modern attachment collection with point-in-time versioning and item-version metadata, enabling accurate sync filtering - Improved API robustness with domain-specific exceptions and correct HTTP status codes for missing resources Tech: Python, REST APIs, enterprise SaaS integrations

LEGAL

AI / Backend Engineer (Python)

LEGAL • Meio período

Jun 2025 - Present • 1 yr 4 mos

Backend/AI engineer on LegalPA, a production AI legal assistant where lawyers chat about their case files and get streamed answers with citations, grounded in law, case law, and specialist knowledge bases. Second-largest contributor to the chat API (~1,160 commits). - Built a single LLM layer that runs GPT-5.x, GPT-4o, and Claude Sonnet on Azure via LangChain, with fast, reasoning, and deep-research modes and an LLM router that picks which jurisdictions and knowledge bases to search - Built the law and case-law search services on Azure AI Search with hybrid retrieval (keyword, vector, and semantic reranking) and LLM-written search queries - Added parallel LLM relevance filtering that removes irrelevant search results before the model answers, keeping responses grounded in real sources - Cut LLM costs with prompt caching for both OpenAI and Anthropic, cheaper models for background tasks, and token budgeting that processes oversized case files chunk by chunk - Built 21 of the platform's 31 data ingestion pipelines (scrapers, section-aware chunkers, embedding and indexing jobs), indexing 11,400+ insolvency cases into 71,500+ searchable passages, with change detection so only updated documents are re-embedded - Added Sentry AI tracing with per-call token tracking and fixed scheduled jobs and retry logic that were silently dropping data or stalling on rate limits - Wrote ~1,460 automated tests across the two codebases Tech: Python, FastAPI, LangChain, OpenAI and Anthropic SDKs, Azure AI Search, Azure OpenAI, Cosmos DB, Docker, Azure Container Apps, pytest