Jul 2023 – Mar 2025 · 1 yr 8 mos · New Delhi, Delhi, India · Remote

• Gandhi Today (A RAG Application)
• 1. Developed a chatbot based on Gandhi's principles to address modern socio-economic issues.
• 2. Leveraged the Collected Works of Mahatma Gandhi, splitting documents into logical chunks that are neither too large nor too small, ensuring sufficient context for effective LLM responses.
• 3. Implemented Cohere Embed English Model to convert chunks into vector embeddings, storing them in ChromaDB for efficient similarity-based querying.
• 4. Designed a pipeline where incoming queries are vectorized using the same model and matched against the stored embeddings, retrieving the top X most relevant chunks.
• 5. Integrated the Ollama Model, providing the retrieved chunks as context alongside the user query to generate precise and document-driven responses.
• 6. Achieved highly accurate and contextually relevant answers through an optimized Retrieval-Augmented Generation (RAG) process.
• Stack: Django, AWS Bedrock, Redis, Google Gemini, Ollama, Cohere Embed English, Chroma DB.
• https://gandhitoday.cloudrunr.co/
• Github:
• https://github.com/hatim009/aws_bedrock_access_gateway
• https://github.com/hatim009/gandhi-today
• Haritika.Org
• Built Backend for a data collector and management app for an NGO, which would help collect data for Farmers, Farms they own, their Beneficiaries and environmental projects that are running on their farms, Also enabled user management with admins having control on who can do what in the app.
• Backend was built in Django a Python Framework, PostreSQL on AWS RDS as Database, AWS S3 for file storage and AWS ElasticBeanStalk/EC2.
• https://github.com/hatim009/haritika_org
• Cloudrunr
• Built a cloud comparison tool (https://app.cloudrunr.co) for Cloudrunr. This helps customer link their AWS account or upload their in-House server usage and compare their existing cost with leading cloud service providers i.e. AWS, Azure and GCP.

Dec 2019 – Mar 2023 · 3 yrs 3 mos · Bengaluru, Karnataka, India

• Worked on Purchase Similarity feature which would remove similar products which customers have already purchased from their email recommendations. For ex- If a customer has purchased a T.V. from LG do not recommend this customer any T.V. from any brand until he is ready to repurchase.
• 1. Enhanced accuracy and expanded already existing solution WW across multiple categories for offline batch processing system.
• 2. Created a High Level Design for a near realtime service (latency < 200 ms) which would tell if the recommended product has already been purchased by customer or not and decide if it should be recommended again. Challenges involved precomputing Product Title to Vector of around 4B products and store them in a key value store. Timely refreshing Product Title to Vector values. Choosing a database which is read heavy and supports batch write and keeping Database read/write cost within limits. Achieved this by removing less popular products based on number of views in last 3 months. hence, reducing number of candidate products by 400MM.
• 3. Successfully Implemented and Integrated the realtime solution in a controlled manner through A/B experimentation and monitoring and expanded WW. Helped remove 5% of marketing emails that were irrelevant to customers.
• Helped deprecate legacy systems and replace with systems which were robust and had less operational overhead
• Mentored juniors

Aug 2017 – Nov 2019 · 2 yrs 3 mos · Bengaluru, Karnataka, India

• Helped migrate marketing email program View Followup Email (VFE) from V1 architecture to V2 which was more scalable, efficient, robust, easy to configure and maintain. This program would process TB's of data to generate candidate emails for around 80 MM customers. Reduced runtime of program from 9 hours (Hadoop) to 1 hour (Spark).
• Developed a Java-Spark based component for VFE that segregates customers based on multiple attributes and conditions fetched from diverse S3 sources. Designed a concise High/Low Level Design document after evaluating approaches to select the most configurable and extensible one. Enabled A/B experimentation for impact analysis. Created a customised solution for US operations to mitigate scaling restrictions by dividing targeting data into multiple segments and enabled automated processing of each segment at different intervals, which helped reduce issues/operations overhead arising due to scale.
• Impact: Increased email sent count by 6% via segregated High/Low Engaging Customer campaigns, boosting program value. Improved email click/open rate by 1.2% while keeping unsubscribe rate steady. Achieved $7 MM annual incremental sales. Processed billions of email notifications in approximately 2 hours

Jul 2015 – Aug 2017 · 2 yrs 1 mo · Greater Delhi Area

• Highway-Social:
• Created a dynamic social app enabling real-time upload, download, and sharing of multimedia content (Video, Audio, Image, GIF).
• Features include Like, Share, Comment, Flag, View Count, Repost, Profile Views, Follow/Unfollow, Category-based Posts, Notifications, Feeds, and Search.
• Challenges overcome: Generating and delivering real-time user feeds, managing user actions (Likes, Reposts, Shares, Followers) efficiently.
• Marketplace-Framework (C26-Store):
• 1. Engineered Backend APIs for Store app and Client Web Panel.
• 2. Achieved server scalability to handle 1000 requests/second load.
• 3. Optimizsed performance via Redis Cache, reducing query redundancy and response time.
• 4. Implemented efficient data storage in ElasticSearch via Kafka for analytics.