Building scalable platforms that drive real impact.

Problem

IndiaMART's Phone Number Masking (PNS) extension solution was significantly underperforming compared to its static number counterpart. Buyers faced an 11.9% call drop rate and an average 11-second delay before a call would even begin ringing. This friction in the buyer-supplier journey directly impacted the marketplace's core value proposition: seamless connectivity.

Constraints

• Existing extension architecture relied on telecom-level DTMF tones, adding 3–5 seconds mandatory overhead per call.
• Zero downtime tolerance during migration of 10M+ supplier accounts.
• Managing "Dual PNS" risk: running two number systems in parallel for 30 days increased operational complexity.

Approach

The strategic insight was that the performance gap wasn't in the extension model itself, but in the telecom-side DTMF processing. The hypothesis: moving "intelligence" in-house via a Reverse Lookup service would bypass this latency. We repurposed high-performance Static DIDs to act as "Virtual Extension DIDs" (4-digit extensions) to get the best of both worlds: speed of static numbers, scalability of extensions.

Execution

I led the architectural pivot to an in-house Reverse Lookup logic that identifies suppliers on the pre-call request, eliminating DTMF read times. We repurposed 1,200 Static DIDs to serve as 4-digit extension DIDs, creating capacity for 10M+ suppliers. A phased rollout began with 45,000 suppliers to validate KPIs, followed by a full rollout over 4 months. A 30-day "Dual PNS" window ensured no leads were lost during transition.

Problem

Desktop users experienced significant friction when contacting suppliers. Unlike mobile users, they had to manually type displayed numbers into their phones. This manual bridge caused a poor 30% click-to-call ratio—a major drop-off from intent to action on high-value B2B leads.

Constraints

• Feature needed careful safeguards to prevent misuse and ensure system stability.
• Needed careful A/B testing to prove "Active" messaging worked better than "Passive" without alienating users.
• Phased rollout required to monitor telephony infrastructure load.

Approach

The core hypothesis: the platform should actively bridge the device gap. Replacing the passive "View Mobile Number" CTA with an active "Call Now" button triggering a double-leg call (system calls buyer, then patches supplier) would reduce friction. We A/B tested variants like "Talk to Supplier" vs. "Request Callback" vs. "Call Now" to find the highest relatability.

Execution

We modified backend services to trigger automated double-leg calls. I led the UX redesign to replace the text reveal with a clear action button. We ran iterative A/B tests to optimize messaging ("Call Now" won). The feature was rolled out in stages (10% → 50% → 100%) with integrated spam detection to ensure system stability.

Problem

Business-critical communications were fragmented across personal WhatsApp accounts. This "shadow communication" caused three risks: operational inefficiency, PII exposure (employee numbers visible to everyone), and zero executive oversight or audit trails.

Constraints

• Strict WhatsApp Business API compliance and template approvals.
• Needed multi-tenancy for different teams (HR, Sales) to share infrastructure but keep data isolated.
• High sensitivity of user data required PII masking from Day 1.

Approach

We needed to "productize" the channel within the ERP. Instead of a 3rd party tool, we built a native integration using the AiSensy API. This ensured multi-tenancy, user access management, and kept the audit trail within our existing BI ecosystem. PII masking was treated as an architectural constraint, not a feature.

Execution

I managed the API orchestration design with the engineering team (React/Node/Go). We built a Permission-Controlled Template Manager so only compliant messages could be sent. I oversaw the asynchronous processing architecture to handle bulk blasts without slowing the ERP. We included webhooks for real-time delivery status.

Problem

High volumes of automated IVR and robocalls were hitting the platform in rapid bursts, severely congesting the network. This degraded the seller experience, overwhelmed operational centers, and prevented genuine buyers from connecting with suppliers.

Constraints

• Thresholds couldn't be too aggressive, risking blocking legitimate power users.
• Detection had to be real-time (sub-second) to catch bursts before damage was done.

Approach

We needed a content-agnostic "speed limit." The strategy was to analyze attempt frequency in micro-intervals (seconds) to isolate bot behavior. We extensively audited call logs to define conservative, safe thresholds (e.g., 3+ attempts in 2 seconds) that exclusively caught non-human behavior.

Execution

We deployed a real-time tracking system to identify and permanently blacklist numbers violating these micro-interval thresholds. We implemented explicit data tagging ('SPAM_IVR') for these blocks to ensure clean analytics. Post-launch monitoring ensured zero false positives for genuine seeds.

Problem

Manual auditing of flagged users was a severe bottleneck. Genuine buyers who were incorrectly flagged spent excessive time in a blocked state—especially on weekends—leading to lost business opportunities (17% call drop ratio due to false flagging).

Constraints

• Human auditors struggled with high daily volumes and linguistic barriers.
• Complete automation risked false positives (banning good users permanently).

Approach

Routine audio auditing is a prime use case for GenAI. We hypothesized that an LLM prompted with call transcripts could match human accuracy. To mitigate risk, we forced the AI to output a "confidence score." Only high-confidence (>0.9) verdicts were automated; edge cases routed back to humans.

Execution

We ran a POC on 400 historical cases using Google Gemini API. After validating accuracy, we integrated Gemini 2.5 flash-lite directly into the moderation workflow. The system now automatically blocks or unblocks users based on high-confidence verdicts, completely bypassing human queues for the majority of cases.