IRFANMOHAMMEDAHMED

problem  Publishing one SEO blog post meant hours of SERP research, competitor analysis, planning, and writing — per post.

built  An autonomous n8n workflow that scrapes live top-ranking pages via DataForSEO, then chains three models — LLaMA 3.1 8B for extraction, GPT-4.1 mini for filtering, GPT-4.1 for strategy and final writing — with human review kept at exactly two checkpoints: the content plan and the final draft.

impact  Actively used at Tericsoft to research, write, and publish blogs live on tericsoft.com — a multi-hour manual process cut to minutes.

problem  Engineering managers at Tericsoft stitched together reports by hand across GitHub, Jira, and Keka HR — hours of cross-platform digging for every question.

built  A four-model agentic pipeline over 35 live BigQuery tables: LightRAG retrieves schema, Gemini 2.5 Flash writes SQL, BigQuery executes, Mistral Large visualises — one conversational turn. SQL is dry-run validated and auto-repaired by Gemini up to five times on failure. A custom LoopAgnosticLock solved an asyncio conflict where LightRAG's module-level locks bound to the wrong Streamlit event loop.

impact  An 11-chart productivity dashboard plus a freeform NL→SQL chat — managers get instant, accurate team insights instead of hours of manual reporting.

problem  Appointment booking needs a human on the phone during clinic hours — and stops the moment they go home.

built  An ElevenLabs conversational agent wired to Cal.com through Make.com webhooks. A check_availability tool returns the nearest open slot in real time; the agent collects name, email, and phone, reads everything back for verbal confirmation, then a book_appointment tool creates the booking. Shipped with a custom branded patient-facing clinic frontend.

impact  24/7 appointment booking with zero humans on the clinic side and zero operational cost per booking.

problem  Post-purchase follow-up calls at a car dealership were manual — so they happened inconsistently or not at all.

built  A pipeline where adding a customer row to Google Sheets instantly triggers a personalised outbound call via ElevenLabs and Twilio — the agent greets the customer by name, references their exact car and purchase date, and runs a structured 3-question satisfaction survey conversationally.

impact  100% follow-up coverage with zero human effort after the sale — every call's transcript, AI summary, duration, and recording URL logged back to Sheets automatically.

problem  TIRADS grading of thyroid nodules in ultrasound depends on fine-grained texture differences that are hard to classify reliably.

built  A bilinear CNN with a dual VGG16 backbone that captures the fine-grained texture features critical for nodule classification, aligned with the clinically established TIRADS scoring system. XML-based ROI extraction and dataset augmentation improved generalisation on limited medical data.

impact  Peer-reviewed and published in the Proceedings of ICICC-2025 by Springer LNNS (ISSN 2367-3389).

problem  Sorting objects by colour is a repetitive manual task that's trivially automatable — with the right sensing and actuation loop.

built  An Arduino Uno system: a TCS3200 colour sensor handles real-time RGB detection and classification, servo motors drive precise multi-axis arm positioning, and an electromagnet grips and releases objects without contact.

impact  A fully autonomous sorting loop — detects, classifies, picks, and places with no manual input once calibrated.

problem  Commercial sim-racing rigs cost more than most people will pay to find out if they love the hobby.

built  A custom peripheral on a repurposed gaming-controller motherboard: a handcrafted steering wheel with paddle shifters and control buttons, plus potentiometer-based pedals delivering analogue throttle, brake, and clutch. Chassis, wiring, and calibration all fabricated from scratch.

impact  A realistic sim-racing experience at a fraction of commercial cost — plug-and-play with any console or PC.