Muhammad Adrees
AI Engineer | LLM Workflows & Agents for B2B SaaS
New to Contra
Muhammad is ready for their next project!
Problem💯
When a small business applies for a loan or cash advance, an underwriter has to answer one question: can this business pay us back? To answer it, they manually review months of bank statements, tax returns, and business documents, calculating cash flow, spotting risks, scoring against the lender's policy, and writing a recommendation.
This takes a skilled human 4-8 hours per file, and lenders process hundreds per week. It's slow, expensive, and two underwriters often reach different conclusions on the same file.
What we built🙌
- A platform that compresses this work from hours into minutes, while keeping a human in control of the final decision.
- The system ingests applications and supporting documents, extracts and validates the underlying financial data, evaluates it against the lender's risk policy, and produces a reviewable recommendation the underwriter can approve, modify, or reject.
- AI handles the heavy lifting, reading documents, summarizing patterns, generating draft assessments.
Outcomes🎉
- Underwriting review time reduced from ~6 hours to under 30 minutes
- 5-10× more files processed per underwriter
- Eliminated manual data entry from bank statements
- Consistent, explainable decisions with full audit trails
Stack🧰
FastAPI, Python, PostgreSQL, Redis, GCP, React, TypeScript, and a mix of LLM and document-AI tooling.
Best fit for🔥
Lenders, fintech startups, credit and risk teams, MCA/lending platforms, and anyone building AI for financial workflows where decisions need to be fast, auditable, and trustworthy.
1
193
AI Document Processing — Turning Messy Financial Documents Into Clean Data
The problem💯
Every financial workflow starts with a stack of messy documents,
bank statements in inconsistent formats, scanned PDFs of varying
quality, multi-page reports with tables that span columns, and data
that comes out of OCR looking nothing like the original.
Teams spend hours manually re-typing transactions, fixing extraction errors, and double-checking numbers before they can do anything useful with the data. It's the slowest, most error-prone step in every back-office workflow.
What I built🙌
- A cloud-native document processing platform that turns these messy
inputs into structured, validated, trustworthy data.
- Users upload financial documents, bank statements, invoices, scanned
reports and the system handles the rest: parsing, normalization,
reconciliation against statement totals, confidence scoring, and
review-ready output.
- AI does the heavy extraction work, deterministic validation catches errors before they reach the user.
Outcomes🎉
- Document processing time reduced from hours to under 5 minutes
- Eliminated manual transaction re-entry from bank statements
- Auto-reconciled statement totals catch extraction errors before data is exported
- Versioned results so users can compare AI outputs and track corrections over time
- Export-ready data that drops cleanly into downstream systems
Stack🧰
FastAPI, Python, GCP, Firestore, document AI tooling, React,
TypeScript.
Best fit for🔥
Fintech teams, accounting platforms, lenders, back-office automation
products, and anyone building workflows where document quality is
the difference between automation that works and automation that
creates more cleanup work than it saves.
1
177
AI Bookkeeping Platform — Books That Maintain Themselves🔥
The problem💯
Small business owners don't want to think about bookkeeping. They want to know if the business is healthy, whether they can afford a new hire, and what their tax bill will look like. But getting answers from their books still requires effort, opening the app, navigating menus, running reports, interpreting numbers, and fixing categorization mistakes along the way.
The tools have improved, but the workflow hasn't really changed, users still drive the software.
What I built🙌
- A bookkeeping platform where the AI agent drives the workflow, not the user.
- Instead of opening a dashboard and clicking through menus, users just ask. "How's revenue this month?" "Did I pay the AWS bill yet?" "Categorize the Amazon charge as Supplies and remember it."
- The agent answers in plain language, takes real actions inside the books, creates rules from corrections, and only asks the user when something genuinely needs their input. The interface becomes the conversation.
Outcomes🙌
- Users get answers without learning the app
- Transactions get categorized and corrections become rules
- Receipts auto-match to transactions
- Reports update continuously, available on request
- Bookkeeping shifts from a weekly chore to a quick chat
Stack🧠
Python, GCP, event-driven architecture, LLM tooling, React, TypeScript.
Best fit for👍
Fintech and SMB finance products, bookkeeping services going AI-native,
accounting platforms exploring agent-first interfaces, or any product
where the next product surface is a conversation, not a dashboard.
0
70
AI Agents & Workflow Automation — LLMs That Take Real Actions, Not Just Chat🔥
The problem💯
Most "AI agents" on the market are still chatbots wearing a costume. They can answer questions, summarize docs, and sound convincing but they can't actually do anything inside a real system. They don't update records. They don't call APIs. They don't make decisions that matter.
When teams try to put them into production workflows, they either hallucinate, break under edge cases, or require so much guard railing that the agent becomes slower than the manual process it was supposed to replace.
What I built🔥
- Production AI agents that operate inside real workflows, taking structured actions, calling internal tools and APIs, working through multi-step decisions, and knowing when to ask a human.
- The agents handle the operational work teams actually want automated:
querying data and answering follow-ups, classifying and routing incoming items, updating records, generating reports on request, running multi-step workflows that combine several tools, and recovering gracefully when something doesn't fit the expected pattern.
Outcomes🙌
- Agents that take real actions, not just generate text
- Tool using LLMs with structured outputs and validation layers
- Multi-step workflows that complete reliably end-to-end
- Audit logs so every agent action is reviewable
- Human escalation built in agents know what they shouldn't decide
- Evaluation suites so improvements can be measured, not guessed at
Stack🧠
Python, LLM orchestration tooling, async pipelines, structured outputs, evaluation frameworks.
Best fit for👍
SaaS products adding in-app AI assistants, ops teams automating internal workflows, fintech and back-office products with repetitive decision work, and any team that's tried building agents and watched them fall apart in production.
0
110